<%BANNER%>

Optimal Design and Analysis of Clonal Forestry Trials Using Simulated Data

xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E20110330_AAAAMB INGEST_TIME 2011-03-30T18:36:42Z PACKAGE UFE0012929_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
FILE SIZE 24419 DFID F20110330_AABRQE ORIGIN DEPOSITOR PATH gezan_s_Page_048.QC.jpg GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5
8593ef2e52d106ebb3c40f9a837e71df
SHA-1
c32e2c3b018027d9df6ce93fe63d7276cb8b5cd3
60694 F20110330_AABRPP gezan_s_Page_099.pro
38834b5cbf30996b0eb50633b0f73e15
a415bc513b52dae6bc092834661a400cec4ddb54
25645 F20110330_AABRPQ gezan_s_Page_093.QC.jpg
b4653b1e6f409d85fad81658ba15e9ae
85853944c93f4f38d1581cd8dfcf7014ecb43446
84256 F20110330_AABRQF gezan_s_Page_083.jpg
6f5245791fce45b7e0fe8efe8426932e
820d6add3a8453d8f4b545a145f9489907dcaee0
74208 F20110330_AABRPR gezan_s_Page_067.jpg
227977eacc95d99a4ada333e750366de
acff79c854eddeb8d2a3aa665b793401672fd3b3
5729 F20110330_AABRQG gezan_s_Page_051thm.jpg
89b4d0a7ea79bf740960dc29345757d3
cbaed0153f90f9d6cd04b2d96b0a26626e7af9f5
68668 F20110330_AABRPS gezan_s_Page_111.jp2
8744bdc7111258e17df1e0f42b3c566e
0f19b084a7e69d8c593f6a0793c5296657b934f7
1571 F20110330_AABRQH gezan_s_Page_075.txt
3438af8d1ce3e670a516f811ec39109e
3dc3c7bbbac3bb95e26d5bc1ed249c81d95d1480
64856 F20110330_AABRPT gezan_s_Page_077.pro
cf4362c319017c66f33e2ffb1bae97bf
3d56671c52045ed377d1d6a35985166f437d7abc
6406 F20110330_AABRQI gezan_s_Page_087thm.jpg
15caa1c12d70cb2b0b8393f84d7ad4df
f0c5849eb04814154faf05604585cf292d115224
1053954 F20110330_AABRPU gezan_s_Page_073.tif
aa5d60fe6f49eb07002926fd0d3dc6f0
8475e1b103bf4799bb0c70d4b711eff59099c471
5960 F20110330_AABRQJ gezan_s_Page_082thm.jpg
089c38c91a68022d7786ae3d21090281
74d820c09e834f19780557fd78e0f8dc2b549d1f
392 F20110330_AABRPV gezan_s_Page_042.txt
b14d0dc93931aec2d6c7bab668b8926b
fa4568af97dbdc6a68324405472440bd5666e439
25271604 F20110330_AABRQK gezan_s_Page_007.tif
93395f7722000a20cccad92905b5e778
72b8ca4155a175a8575bd24554bb477fe0914739
11208 F20110330_AABRPW gezan_s_Page_131.pro
9b444d141dec686806e71699929eed0a
b57afcdb0f0b6fd7ce65f5cae3ca05bb8d057a56
F20110330_AABRRA gezan_s_Page_015.tif
e0bd7435a3eab38d7793f443d17b0487
a32327ff25a86ce308ad4d1b35405042b44acc7f
F20110330_AABRQL gezan_s_Page_116.tif
ae7fb954c3871c9ed12003b18045b5c2
8f9004a6228dcea3499bd24f715393c9ab210f33
111479 F20110330_AABRPX gezan_s_Page_104.jp2
f32673c4094490a694c3c949ea6848a5
0c086e7625825a89ecf438d1d1db01ac6f0bfe78
F20110330_AABRRB gezan_s_Page_016.tif
6a9528d038ad50aa1214e5b103b89a93
43c939f89155f0a71c11bd196bcc9fae06e3b3df
119882 F20110330_AABRQM gezan_s_Page_130.jp2
3ffd699a3f7c54715dac1081cc685083
5ca7e410eee4ed347536e0ec8aef1cfe4277a8b1
6038 F20110330_AABRPY gezan_s_Page_088thm.jpg
9fabf01e19af56427ed2b1ed80b87332
e5e41e5b33691677c76514c958780af09ca18317
F20110330_AABRRC gezan_s_Page_017.tif
c158badc6f1eb2b166776dba7b298da3
4f16e4ed4d0a55be38d0f15fa4a753c8e2f284ca
51921 F20110330_AABRQN gezan_s_Page_020.jp2
a5151c0a7ffe4051087696cfa0c51d31
340434d2339861b8980dd378e2c0235739a63c47
91600 F20110330_AABRPZ gezan_s_Page_130.jpg
22db37fb4612c085e16aa9917c760988
e8862bfa13746d60d938f729c5dce82a1fa14e5d
F20110330_AABRRD gezan_s_Page_022.tif
4f49e32c1a30a693ab76053d81ebd151
7e8adfb0874c42ac47b515fe04e334a44d7fa05a
17425 F20110330_AABRQO gezan_s_Page_106.QC.jpg
ef2a6e7886d3b685615abd100295d5e4
0d8b88f7656232dc27fe4153c958e4f8cbd25b22
F20110330_AABRRE gezan_s_Page_023.tif
520dd1af0855caeb948b4585008bde2f
96de840ea2788cfc4d306e6dbabe67ba0558c902
6451 F20110330_AABRQP gezan_s_Page_131.QC.jpg
d930e2b29d0dba1fbb7f1559987caa5a
acd7e24c2b613d0e272c7256c7bbc726a596ed92
F20110330_AABRRF gezan_s_Page_024.tif
6f221d1e6ab72f48131689cdd7bb738c
63a13a4e31888173993a97ba49d74f57d80896d6
49896 F20110330_AABRQQ gezan_s_Page_082.pro
3eda4ec61a0fbedccc96292f65b84771
757e1a6cb21838303388e0a2afccd2b980d8e4f6
3253 F20110330_AABRQR gezan_s_Page_122thm.jpg
cbe7338544bc3b57208cc14abc32f5ce
2c20e583e287a47667ff9d0f95253fe57ac78325
F20110330_AABRRG gezan_s_Page_025.tif
67dce771fdd81a4267dbb3d4090bdb2e
d41ce923acd027f01e01c195af555161131fd5c5
109808 F20110330_AABRQS gezan_s_Page_084.jp2
319951936bbe26fe4aa221cc4a0e88eb
4b04f09adbaaa070ce66cd9492836629e851ea4c
F20110330_AABRRH gezan_s_Page_027.tif
f1b29b860a23a56d075220d645e5182b
c24882ddffe6fab7ca2d392efad340863b35ce00
212203 F20110330_AABRQT UFE0012929_00001.xml FULL
896bf426a1c8478fad0d9c6058397380
da6d80f6e4111ea6e6ad6f27bfd75771b09b33a3
F20110330_AABRRI gezan_s_Page_028.tif
970fa0ce0e2a211c11dfb1a6bc401c99
dc7c97647e2abfa626aa22bdcf302c8ee843b562
F20110330_AABRRJ gezan_s_Page_029.tif
5420df728e200dcf63dda145d95daa42
a1f49f13c9e50f2ebd854b8a70977dc9170b1a75
F20110330_AABRRK gezan_s_Page_031.tif
5bde58356d64156009cba9943552e3dc
a667f996784b905a928a03f72a71f9c2dc3d93cd
F20110330_AABRQW gezan_s_Page_003.tif
7ba00863cb584446304b619e19d2b784
ba3c1a17fb323419b5d40e6da5d6ef3695231995
F20110330_AABRRL gezan_s_Page_033.tif
96fb263cfbddc64a2bead2ca072902a3
f96752a2f95ec8b33e1b52ea0d6adf03ac91d372
F20110330_AABRQX gezan_s_Page_004.tif
81bf5a22697774970f370366d3ad4a27
4af389cc5f6b4c6f46ae5d64b3ce04b6e853d396
F20110330_AABRSA gezan_s_Page_074.tif
c684c2c2d75783eace67ccdfd92f916e
06caff1f8c7bc6001f989179ee23d7e643f7e7c9
8423998 F20110330_AABRRM gezan_s_Page_035.tif
83371afd544d956ddbe92b49d5a8cd89
ea9e75416c745e57f956e823ceaa27bb9484561d
F20110330_AABRQY gezan_s_Page_011.tif
fa9f4e721c45b0ea88bad073f9a6d48a
df780c48e7f0f66366e9991fd066242fc0256a4d
F20110330_AABRSB gezan_s_Page_079.tif
8a107d785a81ec0136c65eba71cc5b8f
6852800a5dc29a046a5ce7d1a5b2c0eb021912fb
F20110330_AABRRN gezan_s_Page_036.tif
50cfcc47d2986552ce3e3aefa7f97e71
5a6297c5cd0a0010ec291e0206b2c80125ecdbca
F20110330_AABRQZ gezan_s_Page_013.tif
54361e6f9cfb6aa459cc9a639423336c
a4314043fee43cd4adc8d0343facd90a35887757
F20110330_AABRSC gezan_s_Page_082.tif
cd6afe1dd91211d4add1eb086b23e1fa
fd92d16eee7242869126d5277f3b45e4730fd42a
F20110330_AABRRO gezan_s_Page_040.tif
1cebe4afc99c31f8fa924e62b940afb8
1a43f5de4f21230e9fb5d7399a3e4658236dc647
F20110330_AABRSD gezan_s_Page_083.tif
f6887ad05359e1c2da2352d192f3a40d
6c940445104a3dcfc5ad991054970dfa6e2195c4
F20110330_AABRRP gezan_s_Page_041.tif
655adde10101819aef2200955514a171
08481e543e92b04910c4aaaa03f1f9cbfe784b49
F20110330_AABRSE gezan_s_Page_084.tif
f99759fccb570359003c78d20af9f646
8d0c92073c67a1643ed6337922e3b5495812e862
F20110330_AABRRQ gezan_s_Page_047.tif
244b3759dc01be9f2f9a34ffd21bcbbc
3aa9dcffbb08890325256884db6eba6cdc764049
F20110330_AABRSF gezan_s_Page_085.tif
556ccf52d2377a9e577d49375ed2ea52
acf71615bf27909bf534a4519adbc21b85f5ea20
F20110330_AABRRR gezan_s_Page_048.tif
cf8839fd7a18e0953983a3dcfa7d8880
4dc92a28186797b695f4e6261e1a188b8a0961d3
F20110330_AABRSG gezan_s_Page_086.tif
9c42f9b816ff1b0acc5b32906716fc8e
5acbee583711d26a311c7d4fdfa8272096c6f71f
F20110330_AABRRS gezan_s_Page_050.tif
290b087b8cb676e4ab5abdbec3980400
354e0dfd62768db33e9396b690d153283c44b6a6
F20110330_AABRRT gezan_s_Page_052.tif
690d1009b3129321ab3200988738e59d
3066692659dece28002060b544a1f22e9e0eeda4
F20110330_AABRSH gezan_s_Page_088.tif
f446a4a3ab912640f40dc49e85ebd605
4c2a1f3cbca7fa81f5ab0ce99867383ba475b803
F20110330_AABRRU gezan_s_Page_053.tif
7624a873804561a0e76fc8d087f94ed3
db085827995e3f1aeefe166c89e90007e8cacd3a
F20110330_AABRSI gezan_s_Page_089.tif
76eac2b2e0790900cfe9947cc4e5bb8c
fcef5e20e00de50fdf4c3ced9739606d7fea01db
F20110330_AABRSJ gezan_s_Page_091.tif
b3d8c5b257d9586bbe394a6a2118ae7a
3386c7b51139fdbdb53da003b3f1be31aea0d67d
F20110330_AABRRV gezan_s_Page_054.tif
258f523c6feb6fea78e96658249b4c6d
b2f854e1ccbcb6ddfb4850d7c0d383d695deab12
F20110330_AABRSK gezan_s_Page_092.tif
b4b0ea6534a649eed35dc926f81fa085
9a4f9ca2851b8e908c40dbac8e7459719942ac6c
F20110330_AABRRW gezan_s_Page_058.tif
335425eeacd4fc7d43077ff03ff6fca1
f598ab5961a831ec9279838a13d59aabeb5ca52e
F20110330_AABRTA gezan_s_Page_118.tif
d15101cddd6401129ee3fc7775c2b340
c9324cdbbc6e9f67e3acb72f30b2bd2e84f03ade
F20110330_AABRSL gezan_s_Page_095.tif
d8abd3ac132b539aafab9b1f1f49cec8
6977c67141653a63b1c0f7da74cf7d61a15adee6
F20110330_AABRRX gezan_s_Page_061.tif
9cb5fd548c3f7a59e4d0aaff91aa592f
f15dc8b44eabac403ad2e7fc722c3cd54e898cb3
F20110330_AABRTB gezan_s_Page_122.tif
641cd163457ddfe2fc8d31c919622697
a8b6d6d5b4ca9c2bb2d59d579bbbba864ea80357
F20110330_AABRSM gezan_s_Page_097.tif
5c7ad4b5adf7980d698558cabcbdb84f
e2da72d1e13d73260dd09a0da481b49d69bd450f
F20110330_AABRRY gezan_s_Page_063.tif
905209d87c34bb789963defb9f949239
8f4b3fc1787769c91787de6ba8918afe2591c76c
F20110330_AABRTC gezan_s_Page_125.tif
6dde0f3776b1241357c3644e2ffeda56
265fbed584f3d8fdfd498bbf49cebb32b0d2eb62
F20110330_AABRSN gezan_s_Page_099.tif
3125a791be4d03a3d6e99417620a05b8
77d20fc8a2b01e05d9118adeed1e9019bbebcd6f
F20110330_AABRRZ gezan_s_Page_068.tif
556ceb9c963515872ea99a6732bc8490
ac4b550967fcf5f578e81e66de3c8b357a9d458f
F20110330_AABRTD gezan_s_Page_126.tif
63ccc1c3b5e22cc39c21cc7c10bf7f8f
f6b440228f1cd4bbc063c89e0e3a83535c803cc0
F20110330_AABRSO gezan_s_Page_101.tif
f8d016b08099ae3d93df730aca11b0b3
7ffdbd3f49c7a9fa05073a7c896590a88cecad0b
F20110330_AABRTE gezan_s_Page_127.tif
834876c9e0e82d5802e704c6307e1cc4
9766b34e3594c919db8e332bfa55eb15ea3e0b79
F20110330_AABRSP gezan_s_Page_102.tif
435be81f61e5a4dcb73441b7ae8912d3
223832efc238f0628496b0ed99cd0926f655809e
119 F20110330_AABRTF gezan_s_Page_002.txt
e41b03cb3149bd4c09c1010f8bef54eb
fbed6cd643485cafd2de3512a59a548532ea415e
F20110330_AABRSQ gezan_s_Page_103.tif
36a4841d5291196e542718c1db9fc5e9
e4e3c2921ed966c5add8f78eb76f488c85941b35
657 F20110330_AABRTG gezan_s_Page_003.txt
a9985df6a71e37cb9058e1c5e1156745
203c78e56fadbc4964925bb0a83304fd6d91dd2c
F20110330_AABRSR gezan_s_Page_105.tif
07634d5c5fb263499e560f9b6000547b
1e98b821e0095845cd50dbd54fb8e29099a072f5
3082 F20110330_AABRTH gezan_s_Page_005.txt
26f5ab41b05994c510965f863fd82e62
210c33089c2b500ddeac1efadc39904924ecbaa9
F20110330_AABRSS gezan_s_Page_106.tif
c84e3e613f4a6c3e15c5f6c0707d14e8
fd3efd70e32f45ecd4ebeb1f874f7bd13f7a40ac
F20110330_AABRST gezan_s_Page_107.tif
cdb3edb2b8e48de4a8b031e20531736e
4d1bc8dac2bf744b7741e8baa9477670cfbedcdc
2588 F20110330_AABRTI gezan_s_Page_007.txt
a0b5325d8b851f17abcc130bb6e4bf23
ba5083ad857479acb201f532e1e171446fea973a
F20110330_AABRSU gezan_s_Page_108.tif
758fb864ff46dce9393d43dbb1a1e298
bf83ccccc0383b80da420c8b76f04929cc888064
3195 F20110330_AABRTJ gezan_s_Page_008.txt
50393adffb9db33007184cc17f2f5222
b7ff79c617df061b35ea2b77f56033285052f9dc
F20110330_AABRSV gezan_s_Page_109.tif
335efef7e5f39ec1c578ed69f80cc705
73fb70b1f5c60732b19673533d91f11a1dac800d
983 F20110330_AABRTK gezan_s_Page_012.txt
2e66f3e71376053f3ec2882d25be77d7
20ae763c4ccf1b09973c3cde9f48dc443fad06e8
F20110330_AABRSW gezan_s_Page_110.tif
4d53bc7392cc801f29e39370cf6ae997
a7560bde2d8f9d2c3cf73faea03e59e0a3afbc0f
1902 F20110330_AABRTL gezan_s_Page_015.txt
7f0cdf1df5be6b2f1d8d94aabab96f9b
4881c22b3060d1d9af4535ae56a7b63f295ee0df
F20110330_AABRSX gezan_s_Page_111.tif
4c2be553ae9ae257b5cacb584474368d
4691c08b826ed8333c6947feaa4deb63339f915e
1777 F20110330_AABRUA gezan_s_Page_063.txt
98960ae3707c235644ed97f07fc31a36
d7c61b7401af42fb94d651eda8c7cad36e6ec55b
898 F20110330_AABRTM gezan_s_Page_020.txt
1884887fc089f3dd7c6189a3bd019e04
5cdf6b932be242582c5653c09c5cd974f3f2d844
F20110330_AABRSY gezan_s_Page_112.tif
b0322c2d115f7b844e4f7d3f4b46f8c0
6fbfc2b51b720124241fb60bf30885632eadef3c
1939 F20110330_AABRUB gezan_s_Page_065.txt
8f496d55e189c7046c5d3d9e5e2f1bc4
e3b20a9e0b90a25df7cdb48bb3a2e47bb3a5a31e
1784 F20110330_AABRTN gezan_s_Page_021.txt
397228cfedbc2bfd93ee5817976ad9c2
770fe6ba938de5ad95053d2b71de2784885c4e7a
F20110330_AABRSZ gezan_s_Page_117.tif
bb2f224b6526c14e3e301ba963bc5db8
635184acc447cd4188830fc597fa4b308a0acf56
2040 F20110330_AABRUC gezan_s_Page_066.txt
742ecadfe62b5db9589781a9c840b7e3
d2df698bbea8993b2b9a69449e9be281de8cdc0c
2013 F20110330_AABRTO gezan_s_Page_022.txt
55b6bde3bc3d5356ada9f5d9330d9bce
559717ec24d7d30d1068494f2501bb51e8c30f86
1854 F20110330_AABRUD gezan_s_Page_067.txt
58f1fe981082d18b861cc918fbfedfa6
c5737def94ca862baa521d0ea8aba5602755ce0c
2018 F20110330_AABRTP gezan_s_Page_024.txt
d99ce9fe26dc537545ae7d07c56eaeb8
7c7d35d81ea9f11ee056c16ab75a24eac62635a3
2273 F20110330_AABRUE gezan_s_Page_070.txt
4833df0d4973e8bd10ba1cf422675879
e223c60fd7a76720dc21d4bcb94fb15d1a9b918b
1764 F20110330_AABRTQ gezan_s_Page_028.txt
6901582447cb8abe56dc083ac58cb127
62114a3c51b80d3d95e6432a0bad81ba20821423
1802 F20110330_AABRUF gezan_s_Page_071.txt
de2337a07cd00bf9be230993de57c2ec
be68203bb2b35bef0970de90b85080d0461a443a
1935 F20110330_AABRTR gezan_s_Page_030.txt
3a112f21aadb896d32490d4dacdcf72b
2b8c41d475b41719d5f712761eed0c8f731ec063
26069 F20110330_AABSAA gezan_s_Page_079.QC.jpg
7cda2d66bceee42160b3e029c8939e0c
21967ecdc850c172e59dcbd56b32e964ecfb8792
1011 F20110330_AABRUG gezan_s_Page_072.txt
af0e340ff1c7e7bdfa25d84ff432db88
53237825d2e1f5ff3d3e23031a6ed8d45970031b
1891 F20110330_AABRTS gezan_s_Page_031.txt
b6d8b68bbc6070c1e7c3358d82400375
93c774aada534080b7264bc64fa4437c83331ee8
77833 F20110330_AABSAB gezan_s_Page_080.jpg
efd3c6a90eb4d55cec1930b5ea71d176
e97a85f082c8373792abdfad82c396a30857fab8
1848 F20110330_AABRUH gezan_s_Page_074.txt
32dcb00d234e5103e8f4af4e9ca3c9be
20875e5e637aec3ad73052ad91754acf7070184e
1874 F20110330_AABRTT gezan_s_Page_034.txt
020fa7b1e86152510f9d20d8930dffdf
8a492570532a84907f65e68d383851c1c2dc51f9
53602 F20110330_AABSAC gezan_s_Page_081.jpg
c2ba3067405ba0f6c4782112bc351af8
c3892f967a3112609e84a5ddfe71768219a09ced
3394 F20110330_AABRUI gezan_s_Page_077.txt
f486293ab1d99103a6ba9aa33e95c5e8
bb31c53fb13ee649f3d4ac733ee5448d19bc8a82
803 F20110330_AABRTU gezan_s_Page_035.txt
cd76b3817ad2d59e545136e339db6e5b
fef13ad1714e5b7ba67a31ab4c89cb088c5371b2
16871 F20110330_AABSAD gezan_s_Page_081.QC.jpg
745768aadf72cf27402164ac3524516a
d66442a6a489aa146530a38699c7965467098925
2160 F20110330_AABRTV gezan_s_Page_049.txt
47b2f9075f4df2567d57b7ae1a9dd29a
8ea6c463083cf6c5c226b936765ba2e783a3c4c2
23987 F20110330_AABSAE gezan_s_Page_082.QC.jpg
fd3a57d43985770a0bef96f5cb387372
b4ccdf8f9726dbefb4987e70a1609f3d06cee805
3396 F20110330_AABRUJ gezan_s_Page_078.txt
34f2ed9a6c8a7e6d322c104cac933ea8
f37d8dec58c674d54df336457a75c08e1f8157fd
1038 F20110330_AABRTW gezan_s_Page_053.txt
d2a5ddf8dff1ec98edaa574f6e009aef
47aade5609e2bf46e850647d0bb8844746226fb2
26328 F20110330_AABSAF gezan_s_Page_083.QC.jpg
57e33700da2f51efc512c9c24bffa320
6005c8861c5731838162ac5bbc7bf71c8f135818
1885 F20110330_AABRUK gezan_s_Page_079.txt
c5c5e23c2590e07d4f512f93fa1b9240
722329c5263cb44be016faacbd942a27ed899b2c
1229 F20110330_AABRTX gezan_s_Page_056.txt
60d3aeb2391d5411b55c01a5cd888374
c77a7c28c3ac73c1c1f85d74cc05a548aabd1715
80647 F20110330_AABSAG gezan_s_Page_085.jpg
15b9b0585c4733e6bdc7ddadcadb3b8a
0fba075e86f36a176d566ded549113cb0374901b
1216 F20110330_AABRVA gezan_s_Page_111.txt
e6398089a350bb26e3fac54163b07e4f
d67961011680f5c2dd1b0acc96dd517a23fbdfeb
1063 F20110330_AABRUL gezan_s_Page_081.txt
b53464062b7f24ee8f0134f2b1eadd8d
f5f50b3d5e01f82bc7c0b0b2c59e45f9ccd64b21
1962 F20110330_AABRTY gezan_s_Page_057.txt
ba4ddc8c5cb24d10ab2144ba55edc698
55dabf413e11061e1a1813334998378af4b6edca
25615 F20110330_AABSAH gezan_s_Page_085.QC.jpg
0533bc8d3a76a86eff93748524526488
b98b17a76bcfaf2a0346c8feb937fee3214547e4
3056 F20110330_AABRVB gezan_s_Page_112.txt
28c6e2291ee0142d00c42efa6f0f1acc
7a05dcca5e32f10beff4a2f12c0f786f17cb72bf
2024 F20110330_AABRUM gezan_s_Page_083.txt
c90c8c274f8b1322183344689e9e35a2
e677b82082a55232943fe717671f1fe50126cfe5
1812 F20110330_AABRTZ gezan_s_Page_058.txt
25e12e95ba0469dca49d9602a7599c90
bc2d2622d6b569884d1057a3028b4f01c20d9cad
26817 F20110330_AABSAI gezan_s_Page_086.QC.jpg
356307b23fd068195b8aa009b9c5a43e
90b7bdd7d22d047df6b3fdba950e0f7a3cb264e9
2991 F20110330_AABRVC gezan_s_Page_114.txt
91a62352efc1729b4bfd40f243e080e7
0f0521d4ff3dd1cc181e6a4e117f97b7f00802fc
2002 F20110330_AABRUN gezan_s_Page_084.txt
e3b9a6547f8b86cc128dccbb7239d8c2
388423a4cd46664f6b08e4f9a271244edd3c0deb
26762 F20110330_AABSAJ gezan_s_Page_087.QC.jpg
0d7e7ef71651fcd3bb53965e795667c7
c4d269f9031d42a85a18df115dec00b2b2cbf3e4
3080 F20110330_AABRVD gezan_s_Page_115.txt
a307c291492d1ff865c1bdb87ebc4c01
78a5d283f1eb347c52c0f1869eca743cf431ee59
2088 F20110330_AABRUO gezan_s_Page_087.txt
1f2c96d0506ecd8ef27856aac0ec0bd3
1733b7206068b639f4d537219338da569a0112ee
25934 F20110330_AABSAK gezan_s_Page_092.QC.jpg
206631e623097f125c5412e4a72d2612
58dfa28035d84487b3cdea868096d753e8e1c508
1922 F20110330_AABRVE gezan_s_Page_118.txt
eb5cc1b94a06e05dad1f535ac0dda656
5775e0423e6e850a79e31f9e63e146e390c42354
2127 F20110330_AABRUP gezan_s_Page_089.txt
aaf358982a66a396ce4a710ba72fd195
e2cd0d429c2fed6d3c07a867404968fbd3fa2330
78417 F20110330_AABSAL gezan_s_Page_094.jpg
0b1153b9ea25d42355e5bbaaf2fafecc
3b63253f4b3661f7dcc8a4a8214dfea9b65c64de
3158 F20110330_AABRVF gezan_s_Page_119.txt
070ad6f2ed49bd44a8315735a1fa6afe
7fdcaa21d644847adb3a2728f39556834a893c1f
1932 F20110330_AABRUQ gezan_s_Page_090.txt
4e5d3e09a39314f288a841f25f3c72ec
1bcd8189c3091e65a517ef2425cdfa7c8caa25c2
24514 F20110330_AABSAM gezan_s_Page_094.QC.jpg
f2a0312d5768b90fc960b81ce098ff26
8bf23646cb40ee87f8ae54d4f8682518910a0ac7
1580 F20110330_AABRVG gezan_s_Page_122.txt
f4493ca83bfb83523435228760058c71
47c8932fce8722d83aade13819d5a6bbfc0513c4
1704 F20110330_AABRUR gezan_s_Page_091.txt
2cc3a8a7ebb3a6042b81b6f697a489be
4575c943067eb4e6a94d7930b21a1668d261fd73
24851 F20110330_AABSBA gezan_s_Page_109.QC.jpg
f91d2d70a3dee25fec914258be09b9c6
d451ddaab4c33b46177a712304d361a472350544
70145 F20110330_AABSAN gezan_s_Page_095.jpg
c1a08d0ca028e811e90831dbeaef9433
806a862d37c355116d507cb09d246c407426cc2e
1957 F20110330_AABRVH gezan_s_Page_124.txt
cae797f639b713736068eea6736be56a
aa8281bf263071c2744dea994477cc60e79a8c28
F20110330_AABRUS gezan_s_Page_092.txt
6ad56b47c26c768ff85a7562c1de883d
3bb8c8092c3ee77d02a7ab1bd4f41b4970b6193d
75310 F20110330_AABSBB gezan_s_Page_110.jpg
9660e4bdd1ff4af9a7c25b250e7d1205
874b4cabf29c499b64ae0eba6fde85f1a3935808
80763 F20110330_AABSAO gezan_s_Page_096.jpg
1b27e0924358ad3beb1482d68daa4390
86aa26ac08697f003230ac8e058d199e44c2c198
2357 F20110330_AABRVI gezan_s_Page_130.txt
81f6c2a1dcb7569285e69a6a929c2408
0085edcb5abb3b1fdd99ae52ffbc0aa2d3508fe4
1825 F20110330_AABRUT gezan_s_Page_094.txt
2b61bb18451f2eda638837f885faca92
c1d77398b04ffd44c8a3eb3032465512446d3101
23656 F20110330_AABSBC gezan_s_Page_110.QC.jpg
a34f109d13dc9c027013a06356fa62be
3e15bd12f9a40de5193e801e5e4a22ecc29d3797
52016 F20110330_AABSBD gezan_s_Page_111.jpg
ad133729c656491985ea70afe1750fd3
51a1ebf791b6f0314d095b1bba448585a0b48c69
25256 F20110330_AABSAP gezan_s_Page_096.QC.jpg
6b1dd536b8d2213414766e55dffc5034
449375afda6999afa2bc96af7504997d1788cf09
12001 F20110330_AABRVJ gezan_s_Page_003.pro
628fbaa4bbe46cdb3cc1c4e88f52e237
e06c3612e4563d895365dabb9a459eef0b7e5345
1908 F20110330_AABRUU gezan_s_Page_096.txt
895da60ac748a239d40a02672009a69e
3a88554cb28a04d147b7f13d501f82c7219d5252
68607 F20110330_AABSBE gezan_s_Page_113.jpg
43587cf291aab86258c5ba7bebb311ab
eb368f54297cd78ba3bb3668406d11dd916a8e58
58479 F20110330_AABSAQ gezan_s_Page_098.jpg
4fb2d7df962b21821c62748bb26bd994
edce28137587ec5f3a79d99f1b677fcd28371475
1806 F20110330_AABRUV gezan_s_Page_098.txt
564a4f309307685725159db559f20430
790db86133c2c2099e3f3d7c3ade6b19a14700ad
63979 F20110330_AABSBF gezan_s_Page_114.jpg
6ce816f86765a3880b8e1eb3bfb25329
5c42baea0fba14b2cc6fae7695ab1397459c48b8
74097 F20110330_AABSAR gezan_s_Page_099.jpg
6f95cb3d8b2d38de52f6c50fc2c8cfd8
a470355ad63ee58724c3d3488899b81701523e74
36265 F20110330_AABRVK gezan_s_Page_004.pro
7660c00da984c9073d226a292ab5162e
05ba97c25737230fbf974b82446d55bf220fdf2b
1997 F20110330_AABRUW gezan_s_Page_100.txt
56af3f0fa2a264fc155772f7ca19be19
1faab20176f58a2b9119f4c01dca470cd7e7857b
68754 F20110330_AABSBG gezan_s_Page_115.jpg
c5a9a1ab9f2ab13a3bdabf6ebc00805d
4657f6af6b51f40940defa66ed971d77d0f47cb3
46096 F20110330_AABRWA gezan_s_Page_037.pro
47164a4e6997e35f6cd72268ff970f8c
339bb892a55a9fb4fb1ef9214becd3d8877ba255
22846 F20110330_AABSAS gezan_s_Page_099.QC.jpg
ede61b1f900ed7e4ff4ca2ae446d7300
b480ef24b71401650313d0c0410a2dd3644bda27
76207 F20110330_AABRVL gezan_s_Page_005.pro
4363bb4a8348257d0511030e08ae049c
f7372559b4b9a316aeb91cd053c4acef6139fc25
2659 F20110330_AABRUX gezan_s_Page_101.txt
5657db24719ab5654587c43d4250cc2b
bc8899e37621598323149e9f0021172f40cf5396
79181 F20110330_AABSBH gezan_s_Page_116.jpg
5b20acd2117179c81d83ed336f361cc7
7ae64ce7022170bf43188caeec4b6c91769facbe
46486 F20110330_AABRWB gezan_s_Page_038.pro
0d007f4369ab8d1131de793665b1044e
e48b18fbbc3e2765c4443e302abf4392722a8599
21599 F20110330_AABSAT gezan_s_Page_100.QC.jpg
170182bc2b9c77fb21b8fdbaa5623680
be07811e1195daf0bac965c3fffe5b2c359d683e
83330 F20110330_AABRVM gezan_s_Page_006.pro
6aae6e38855f62534ca79ca4467667fb
ac64659cf006bdff2b1df1aa1b952a677a0c9af1
2060 F20110330_AABRUY gezan_s_Page_102.txt
c61b4a1efc378a0e8d7ea4342f6c57da
53d54d83d4ee946abf088303987169fd1ed3e2a8
80301 F20110330_AABSBI gezan_s_Page_117.jpg
247e1a3d9929dd4479c0296717e988eb
684e15e09b568ebf2b339c0afe3bca65467a70fb
49666 F20110330_AABRWC gezan_s_Page_040.pro
8be67934329d4338a258eeaaa66d7640
a40a3fcf8eb6734535ce74b8ef5fcb659b86f724
63254 F20110330_AABRVN gezan_s_Page_007.pro
e320d63b5805dff6941856c41b02fc2d
e72301cdf9efd4563e566eebfd8ce6743b99238e
F20110330_AABRUZ gezan_s_Page_105.txt
51c5e8aa761aaadd8219341009c97bf2
30c11a89d02b272d4e1e4d4893f6f320baad5d56
21541 F20110330_AABSBJ gezan_s_Page_121.QC.jpg
fddf764f6eb3e79f87a7c8e7fac101f6
b7665e03c655cf42812e3782ddd632d91c9f8df2
46901 F20110330_AABRWD gezan_s_Page_041.pro
19aea8b12b22a87ff33cbecfe84aa130
e96b34ff1f797ecd14ada722252648b04b813d46
26515 F20110330_AABSAU gezan_s_Page_102.QC.jpg
32bee60ff971f1408f49b6f50f0f14b5
b32d1d6b111625fc824481469f14bae60ed4e07b
66050 F20110330_AABRVO gezan_s_Page_010.pro
17c8020e86a4e8ec65e6437cadd9fa6e
21c5766980c73f0973e5597111a42e6afd3b7b80
48034 F20110330_AABSBK gezan_s_Page_122.jpg
6a1e90b446ce822cff852127665305c8
82abcb9d23aeadeee80798b984c1bfffd3233f4a
49727 F20110330_AABRWE gezan_s_Page_044.pro
c21ed75b42592611b861e4e3d0c8d093
f5a3e60f27f0a31bef28aa29996fa214810bab18
17563 F20110330_AABSAV gezan_s_Page_103.QC.jpg
836361ff9c39eaee10dda27ed0e10df5
b9d6bb4251d5f06a69a03423c5f337ae40c4ac62
78400 F20110330_AABRVP gezan_s_Page_011.pro
bebbbcb841b3a6cacc2f4ea2e08143cb
6a02687a6192e5965d5eff84499c1df4ad5c431a
13473 F20110330_AABSBL gezan_s_Page_122.QC.jpg
6bef2be6d267175349c21297fe02957e
437f3fc5b6ab5ba8f128c5e8df174866919b11f9
49380 F20110330_AABRWF gezan_s_Page_046.pro
48e53ecd930a8c341e1a3fd430594e2a
87d953becb23e9a9fe132d5449ae0b8c45d34043
85424 F20110330_AABSAW gezan_s_Page_104.jpg
bf14fea7a4c1f96ce2540630990c803c
47544df9a056bf671811ed2bef27257c5a336ca3
38497 F20110330_AABRVQ gezan_s_Page_013.pro
38c559f3414fadcbb4fcdee0e60cd183
8cc68bed36b214c8b85f4ed15304134c071d85bb
1051964 F20110330_AABSCA gezan_s_Page_010.jp2
9182f3416e5218a82f195ddbc5366232
c1455781a81d7b394d470e3bd489bbcf65687010
18435 F20110330_AABSBM gezan_s_Page_123.QC.jpg
c7ebb5c4c91f1da12058266f2dee175a
48a160c601220bb2e7b67dae46432eae27902e68
55901 F20110330_AABRWG gezan_s_Page_047.pro
c7317d4e9b92bb797e59ec0365d782c9
0cf316be31438b895d5d56200751eda7df7b66fe
80074 F20110330_AABSAX gezan_s_Page_105.jpg
42cdc44188e350df2d1b24dea05ee998
6b34f6d34c15e70d1aa22189eda7ce5e8c514af4
45218 F20110330_AABRVR gezan_s_Page_016.pro
b288953b287d3337757d0967dccff899
77bda5c3a744c269ee5f63baf614677257939da4
1051888 F20110330_AABSCB gezan_s_Page_012.jp2
07c19965800205c80112332849648c82
e602eb8656d662def99402fa98decd5aa0a2a39a
74795 F20110330_AABSBN gezan_s_Page_126.jpg
4b575c626abc380c2f46c6e352c3ab87
db4894c0d86b34108a55fb346b9bf62296094ab5
20405 F20110330_AABRWH gezan_s_Page_052.pro
fb6d639ac438adc8e9e121c900863314
e6c7fefbe1d0f9a346059401e440bf3b2a5ded16
20503 F20110330_AABSAY gezan_s_Page_107.QC.jpg
f38629aa7b633f1be58144249695ae38
7b1229972e8e2603233afebafc0ad5ab583aa192
48564 F20110330_AABRVS gezan_s_Page_018.pro
3a2915ee9ad23129d3c1a4a5786d8eb4
d1d4b44d8ef450ee6279618e6b06cc2a4755f58b
100292 F20110330_AABSCC gezan_s_Page_015.jp2
7401b6357cdf279bd14f022ff16428fe
0398e2b7ed053a9d4813c66d469ed8aca6d7921a
21354 F20110330_AABSBO gezan_s_Page_126.QC.jpg
d59dcdef51bd7d681a4b32435fa898e9
40f0168352358be70a849f58c11814cdd7242a13
20526 F20110330_AABRWI gezan_s_Page_053.pro
eecb77a4f0387a04ee1316b36215112d
c50c5fcae359bf49c5f252dc43788cfeec51b2e9
24872 F20110330_AABSAZ gezan_s_Page_108.QC.jpg
6c5d162b9e2051cdc954a4494473b17b
705e00251c76c0982846d7255345499d7b30f447
22469 F20110330_AABRVT gezan_s_Page_020.pro
31f25d9fe95b8a3162528f9954ceca3a
603c2f952966ee7ecaea56561187a6ab558939af
1051967 F20110330_AABSCD gezan_s_Page_018.jp2
237df584bc703cb9a264ccfb2f156807
a71c89bf72ffb077b1028eb44d4446d5a4a5d0f2
28492 F20110330_AABSBP gezan_s_Page_127.QC.jpg
cd8bab3c25ec7bc216b62664602b261f
fea8941f6220edd3de3939c82b2625af47402cd8
47275 F20110330_AABRWJ gezan_s_Page_054.pro
0b136e9eea4d59a009453b67b1c52c26
d201c979bf16694e42407d6d251d9580d22f5d46
31513 F20110330_AABRVU gezan_s_Page_025.pro
fa22ef23e68d14a43e0e9d420a9742c3
082551b52343691bab3acf87ebf852120548f809
92446 F20110330_AABSCE gezan_s_Page_021.jp2
a76c42d383640803444564fc8fdb25ab
7dd2e57e4dec2062b686eddf97f55dae7940b8ee
96761 F20110330_AABSBQ gezan_s_Page_128.jpg
4994744446cbbdea4c75dfcd6f3d2367
cd32b0f395d73dbf7905b137c69d00cea7e127e3
30559 F20110330_AABRWK gezan_s_Page_056.pro
f0bf820237f6c094af827fe52ed482d9
434195bce5a7b00738b57a61255da6d4d0d534c9
46827 F20110330_AABRVV gezan_s_Page_026.pro
f725e4fda065cde985779be1adabb6dd
2fc06c63bff142bb1edd5987c21515aa83d62074
110442 F20110330_AABSCF gezan_s_Page_022.jp2
b9ee459e03b78b74b8373bd066917846
5de07021c310186b496d24928d81dab5b6ac3d7a
26503 F20110330_AABSBR gezan_s_Page_128.QC.jpg
04415af9a6bd837ff9d0b95a28b0b20a
6559d9c15054a6f31a639538ee9a30841b6b6210
48598 F20110330_AABRVW gezan_s_Page_029.pro
9348afb839f1fcccee20723f7357e372
d5d161f9d8b503a9fda2d720634a25d9b5494d0e
109733 F20110330_AABSCG gezan_s_Page_023.jp2
9e1ea9053cdd0d945fe1ae44fd599e4c
8e2279e295c778025146dbd453f2684b1f0f75e4
95972 F20110330_AABSBS gezan_s_Page_129.jpg
76057effa92e2a96d4c2fb142a5b0f69
de7cc71d6b2410fdf192bc41cafe4e03f7c563ae
39163 F20110330_AABRWL gezan_s_Page_062.pro
23549b043463cf5e4c5cd55c817fd571
0a363c3754b1b906cbe48a2d11c26dc060855534
49114 F20110330_AABRVX gezan_s_Page_030.pro
80a91250a3e74164aaee044d5b14318d
36adea9bb9aa1877b64a851efa54a51dd0456734
46126 F20110330_AABRXA gezan_s_Page_094.pro
3edacc8c4401140d0037ce418cc3a952
c92eb280b76df59ed60567e79aa8ed4a70fdc5a0
707142 F20110330_AABSCH gezan_s_Page_025.jp2
f73dffb09cb7f8fcb31187adc58e2f42
99fa7b01804d74df62a72af0a7e656237e679ce0
26253 F20110330_AABSBT gezan_s_Page_130.QC.jpg
18952fe7dd58215e4c34c9fbef5c4a55
33ab6e16940f70d10a424482b417c7c2a8845946
45285 F20110330_AABRWM gezan_s_Page_067.pro
72aba8cc1389a6071204e1b6d47ecfa6
06c6070f671fd04eaa1e766d4a8c6aefe0c2eaaf
48281 F20110330_AABRVY gezan_s_Page_033.pro
15431da5d81f7fa51627a2080b7d847a
784425a020087a1dd52282b25dd4a4d90cfa72db
48164 F20110330_AABRXB gezan_s_Page_096.pro
e8906f6b2e6caf34f564b0173ec75b4f
af136f3ba7cb29ff873e50f00cf57745b530092f
75328 F20110330_AABSCI gezan_s_Page_026.jp2
11dc055426ded736998857d92ecca4cb
a10a57e2659285f19cd9867fbcd639af7ed9d969
6413 F20110330_AABSBU gezan_s_Page_002.jp2
6b8b565894b3868f76b08fdebf095be3
92c170354e8480f5d55c5007b51fca715cffcdb7
48365 F20110330_AABRWN gezan_s_Page_068.pro
2391b310f75d3f34153744772dfde76d
a21abde17acb050d5288c55332c33dbf0c8fd56c
19533 F20110330_AABRVZ gezan_s_Page_035.pro
b575b6e52569984454d7d0bee15baadb
e1965b5b102f010892815fb0f116f3488f84ee2b
47292 F20110330_AABRXC gezan_s_Page_097.pro
183b0106bb5e48d7f509bae6672a6895
325f0de20671c61db1304ae334de29cdca8b1b27
104369 F20110330_AABSCJ gezan_s_Page_027.jp2
3d9744d76b6f2a09fc2727384996ac23
784cf2a6c3ee08ebc923215de4c14b3263418899
48146 F20110330_AABRWO gezan_s_Page_069.pro
79eb2b2faa1b885d94457e0db25b4454
0061e21036803cad4ff89d80f6c55ffc75c56348
34708 F20110330_AABRXD gezan_s_Page_098.pro
6f961e2a3f6dab32d272f8618ed89d9f
6221d6550487d988e6963742cbca017d74026c33
942331 F20110330_AABSCK gezan_s_Page_028.jp2
06f4a642aced91d7e4e2cbcc5263eb53
5cf95bb02f6ac0f54a8d199fba74bb6acc87bf52
28211 F20110330_AABSBV gezan_s_Page_003.jp2
7e4ee49e715fe28c4a9b2e35b891d0cb
21acaf30368fa176eb97bdae87f834c55da3d9f7
36457 F20110330_AABRWP gezan_s_Page_075.pro
8fd57ec07cf2c12849267ace1fda88af
359ccf67afc8862f046f5c3473833c648635ae21
48318 F20110330_AABRXE gezan_s_Page_101.pro
b2b64a801945bbea52f4f01346867aad
e63ebc9f1f7f2c95ae35be3b033b68f7a12bd1dd
105357 F20110330_AABSCL gezan_s_Page_029.jp2
8daee2ed57af2d69b1f30b7a323efe07
dec626e7d3b0ef14e2db6b24ee759d2e9fa6d4eb
79416 F20110330_AABSBW gezan_s_Page_004.jp2
1a41edbddab925960df23bc4301bbab1
8622eb6b178e617fd47ee250e9d9f3f35e18d216
47717 F20110330_AABRWQ gezan_s_Page_079.pro
be801b715dd77b3964ca30fc1341aaae
7c610d788138305ffa7dee3d5afe3036eb203fd7
51369 F20110330_AABRXF gezan_s_Page_102.pro
40bcce88f3d4f49d06a8592082e6a4be
5ca4c20a208f2c5f94f5cb328deb6478d60b4c24
425252 F20110330_AABSDA gezan_s_Page_053.jp2
26168b253a71659224df4ffb7a91e9b5
fc0805d33d355ef700e813fc09411bc56b2bc708
107115 F20110330_AABSCM gezan_s_Page_030.jp2
23a1d23fef82821331d0a414df4a1448
81a4997a5167ace116ae62be9bd5ba98c94f0dc8
1051955 F20110330_AABSBX gezan_s_Page_005.jp2
a86d382849327429b82c907f2ff904ec
ed7016c05203f8854c135d28baf9f9548f3aad18
45864 F20110330_AABRWR gezan_s_Page_080.pro
8c2905eb44931b7c86bca93769504f91
1ac8c8b0a830126189d4b5d58cf5763dc097ef2a
48799 F20110330_AABRXG gezan_s_Page_105.pro
2318b4c70f87b34169317b719a0fa8a4
e89af86fa0624751ff38d192bfd3f56db751bbea
112334 F20110330_AABSDB gezan_s_Page_055.jp2
efd384768cb082ba18e84281a04cc7d9
884c26c0fba1cd643134e0c359fdcf1b319678b4
95562 F20110330_AABSCN gezan_s_Page_031.jp2
664880c901bb00f9d11f2c4d2ddd2022
5af9df314c0d102d8384b5ffcdde9917f35b0002
1051980 F20110330_AABSBY gezan_s_Page_006.jp2
12f6b3dd7d60675799db9b069f40d7a0
553f02acabd0efe9f3e536ef92b6b1de69ffea32
47986 F20110330_AABRXH gezan_s_Page_109.pro
574d43200aea0302f5c68a9f4b4f48cd
5f0ea656e0ee9cc53f9b6abada0f2da06e5db3f4
20454 F20110330_AABRWS gezan_s_Page_081.pro
02059da7e441cd7552cbd49952e86b1c
2906fb047bbc2be7259363205ebb922d8860d0da
1051982 F20110330_AABSDC gezan_s_Page_057.jp2
fdedf82582a780ad865b5ba777109272
2409e9ccd1422b75bf2d31ef37cb1be93ea032a4
70186 F20110330_AABSCO gezan_s_Page_032.jp2
994f38ec7fb95e8fd7eb8d2148bb4d38
10f3c67739f179e2a6047977351f082115d4d6af
1051951 F20110330_AABSBZ gezan_s_Page_007.jp2
361affe03d1ee2d5470b823dd391cc89
34120eeb28f25377eca4b3e792b1ddfd1aa573a3
43388 F20110330_AABRAA gezan_s_Page_028.pro
95111b0aff7c9765773a81d8d56756d8
ea5839fe5847268d7871b6e54084dbc9d0a84f1b
54225 F20110330_AABRXI gezan_s_Page_112.pro
4cff2868d8bb13641eafa18db52049d9
273bb863831ca5146b825c60645530997b246e19
48993 F20110330_AABRWT gezan_s_Page_085.pro
827b5a3a0bb21795dd17f5a3518e0c63
3f61c0f4c8d21f690671b8cd7be69d6152a5d7f5
732190 F20110330_AABSDD gezan_s_Page_060.jp2
76f7b571cbae58f26ee1270261a620e4
4273d887530915f25a5dd8d43a2234be11ca05a5
104608 F20110330_AABSCP gezan_s_Page_033.jp2
bc152c315bc833b6016337853f285f37
e741865916d564bb468e3f56315ff1ddbc224503
86992 F20110330_AABRAB gezan_s_Page_064.jpg
75e525b72c2e080130211fa426ec40cc
a0b1096c3d29b3cf7889779a6dc8ac779a4b248b
56912 F20110330_AABRXJ gezan_s_Page_116.pro
6b9b6542112dd32fc653e9dea3546cfd
bb5e97af789449b82ad5df19577a25d4b0fb7f02
50545 F20110330_AABRWU gezan_s_Page_086.pro
6e7124ea5639bd991499213ea375f2b5
c86bc16318a33f55a04d0d49a7a3f961375819fd
101436 F20110330_AABSDE gezan_s_Page_061.jp2
fdcf25cff003ecca74397a413ba44c78
0fa06c5f8ae82acf34d3951cf390152163043a9f
110632 F20110330_AABSCQ gezan_s_Page_036.jp2
957ea60d3e5b5f62dd95c053aab33dcb
d33189e640be94834b7dc2d470741efff38f5a39
1765 F20110330_AABRAC gezan_s_Page_043.txt
9eccd514769228c32680058a8e317c92
27afcc2ec4228fb4c332852045d71896501a33dc
57876 F20110330_AABRXK gezan_s_Page_119.pro
3a506e91475e35b001a793d32f439992
dbe36a7bfe9517187e1d57b2d3c820aa78992bf2
53276 F20110330_AABRWV gezan_s_Page_087.pro
a642969ba09624f9cfb30496aa80d49b
018c1355779ca7e414e344b430fb3b44d2a43648
85836 F20110330_AABSDF gezan_s_Page_062.jp2
855cd8550b51fcd6e3062f04b2a51501
19d7442ee1284c3cd0033165d8eeec019c561d2d
101708 F20110330_AABSCR gezan_s_Page_038.jp2
a5fa6c9c12bd9dd6042382c28925bc79
74cc5f24e0354fcedb9a72edb48da19209fdfa6a
22987 F20110330_AABRAD gezan_s_Page_117.QC.jpg
3535af777836b2aae8c46e8d39ba6ec1
2b896cf3d9d4ca3381b2b3ef96aa7c855cc3d14a
31336 F20110330_AABRXL gezan_s_Page_120.pro
27096bb4e210df574b73f16862cfe5a6
0a4a398a1a0379bb8b36379169002d8d95e9f206
48991 F20110330_AABRWW gezan_s_Page_088.pro
0302d9ebbfb8c01a7f60b8d048746c31
fa05254899648980aafc37a72736ce5dbfe512a9
94066 F20110330_AABSDG gezan_s_Page_063.jp2
21dc15c6af4a5226ae2eb21902e02ced
773e39ed4ca366bced285efd79c4dfe077fed691
684085 F20110330_AABSCS gezan_s_Page_039.jp2
7e08fbc489eee251ef795f65d6a508aa
9875305a45d23e7df3cb6843bba36607b24758fb
79059 F20110330_AABRAE gezan_s_Page_114.jp2
7c0abce328bd3aad5a31fd93aa1c0c3a
bb1ddf5df09b8c3f37e8c29267c9d520aafb0a04
66158 F20110330_AABRYA gezan_s_Page_014.jpg
d626bfff6d18b22ae55aa2a0f2771c98
ce45df300beac05f676f5b3a6047fc04a42b6f53
40167 F20110330_AABRWX gezan_s_Page_091.pro
f1cdf6ae1c8a40d9e05b6fe14255f3d9
fc5c6c00c87af65a3919aa7b613ee00efb5a2828
1051979 F20110330_AABSDH gezan_s_Page_065.jp2
ce9bed990e8bc63920535a456763c5eb
48769149db1095d54e6a970e182dfe2de4fc73e6
108426 F20110330_AABSCT gezan_s_Page_040.jp2
69cc87ce1cbaa3b9a1917cdfed0b9b6c
aa5792adf28da1a5b1b5009af02abf44c6bdae41
20599 F20110330_AABRYB gezan_s_Page_014.QC.jpg
59776d253802492d76160c2faff9f05e
4b93e5f9f0757821834e4574217047380daa277b
23550 F20110330_AABRXM gezan_s_Page_121.pro
3ddbadd143752643a8b73a97c13f4db7
d5adc19fb25871a5326e409342300f83f360db98
51283 F20110330_AABRWY gezan_s_Page_092.pro
c99466f548d44e831b13013b9c9ee080
84f9edf02f41f726820ea492abe90438af0f5312
5796 F20110330_AABRAF gezan_s_Page_047thm.jpg
39745e94fa60cb2edfbf63e1d8f49d4e
d67065c8b767c32139d160a50c638034ab345711
103028 F20110330_AABSDI gezan_s_Page_068.jp2
41122fd1afeddc5db43ca80dd2a6610f
2df0999503909740f22e464d3907e4c1f4587056
104090 F20110330_AABSCU gezan_s_Page_046.jp2
7073b83e03de4f955b8fdd0f94ce8c09
7eb24500cc882df5369468b7e844a2c32971a4c8
23494 F20110330_AABRYC gezan_s_Page_015.QC.jpg
0315b490f25ec2ead18267adcbba29dd
0bdb417b5e9bfae72f0070f80fa1759c700c0c50
34726 F20110330_AABRXN gezan_s_Page_122.pro
944432df3eae61e32c4ca407e3d7c949
7d068efc87f375397fee29eced15acbea6f21963
49397 F20110330_AABRWZ gezan_s_Page_093.pro
3c93904ca31a13214e8784918ede2668
e66ceca2248704bb0fc1c1ea17edc678c118e9ae
47375 F20110330_AABRAG gezan_s_Page_045.pro
befc7d7b380ffa95b504e5be8ce9746c
6761b62ee7ff27cb7da999ebcc458457cb5f3cb1
1051975 F20110330_AABSDJ gezan_s_Page_069.jp2
2ce98d26be10e21ee7bdcd8a7a511d77
03f17084947d4cc98a7b5934539d37d6efbcebc9
100571 F20110330_AABSCV gezan_s_Page_047.jp2
d4f580d77c8ee8eca87f1a50dfb7a7b4
2880c86fb08a4254204c7eeea93c27165ce755fa
85669 F20110330_AABRYD gezan_s_Page_017.jpg
bb9b15760df1759da231800e32b09a82
54ced07633720c5e269f81b9c885d966447ea969
37630 F20110330_AABRXO gezan_s_Page_124.pro
5674bc181563a589bb5759cb6c61b693
d6770b9ed410fd2ef12484458de4d84368dd2c41
54951 F20110330_AABRAH gezan_s_Page_113.pro
5955080840c2db999f5aaf19eb2fd209
5ddf02f4ed4c165f2235b47304d46266d6aa04dd
103605 F20110330_AABSDK gezan_s_Page_073.jp2
b1c2c404c33a55ccf242450730f9735b
5c7665f8128cc78441a0fbcda2f366104ab17df2
26488 F20110330_AABRYE gezan_s_Page_017.QC.jpg
01b8a34ec4c0866c97782f66e4ae64b9
aa72cf718e8ff42e231d3bcc3d7628ede53d3375
62670 F20110330_AABRXP gezan_s_Page_127.pro
fb3aaccd5eee4836761e50e93311bc1b
85b5ade16b153c59cb128c5ef5c4e3177218f031
26672 F20110330_AABRAI gezan_s_Page_104.QC.jpg
0410737b461403ab9fbe0bf1eef0016f
40f9eb1268a8df44df7dc2df789d12baf4b8a02b
92730 F20110330_AABSDL gezan_s_Page_076.jp2
81ca6c370ab41f27099644061a45aea9
e8e179be99230d97b4f95aeaa72a3d755ace6525
102660 F20110330_AABSCW gezan_s_Page_048.jp2
f7326bc04d317789d88c00fe0d34b85e
6909401d34854c92868a94073a99c2a6007bf764
24919 F20110330_AABRYF gezan_s_Page_018.QC.jpg
2e4053632956b559774721396e1253a4
4dca3d4fdbe4adebb21b4ab126096281a3e12ceb
35507 F20110330_AABRXQ gezan_s_Page_132.pro
ede2ae9ba032024d142994c62b2afa8a
cf42ff44a726f04cdb3578d389d45c7660763c93
6432 F20110330_AABRAJ gezan_s_Page_064thm.jpg
40212c13ba3d5e2c5d65f4215be65f45
6e4d2e8c21801db9aa84e135a6dbe767590d7b16
106363 F20110330_AABSEA gezan_s_Page_119.jp2
1e77787b88f7c52a78689635ecfb0b1c
e8a2e9149ce4a0bc7cc584e07b2120b06ce00a03
91861 F20110330_AABSDM gezan_s_Page_082.jp2
412a409deb615163a2f50454f8c65d30
4c3c6ef933df42f32e4e8e0737b638dc590d9e20
108118 F20110330_AABSCX gezan_s_Page_049.jp2
585d1f002536d9ced298981086369161
8565462a46cd399d0d12add3493b22aaf94b48e5
39821 F20110330_AABRYG gezan_s_Page_020.jpg
94634ed6f28590c9d5e694d5c3924ac5
a34602df7745c0e60a04b1bf65091bc8ec536f5f
21647 F20110330_AABRXR gezan_s_Page_001.jpg
f0678b13b98ad169e65c78ed56c1c7f6
3f0108d205974f8bfdec733b891e30996855b458
6389 F20110330_AABRAK gezan_s_Page_022thm.jpg
7917cb96d857fe402ac1d7416f15725d
35d67ff3cd28482b1757bf7032292b269817027f
72408 F20110330_AABSEB gezan_s_Page_120.jp2
1e9d836b8195cb6ba1be890f3159bec8
7f86785e5563743f667f2e22ea4e11a1374917a4
109793 F20110330_AABSDN gezan_s_Page_086.jp2
a8a030fa409b5a3296a0eb4377b73667
5455c480f097fea51457b2c762c4b65afce9a8ff
109095 F20110330_AABSCY gezan_s_Page_050.jp2
8b70d2e308776a1f4e66796d8ca88dd3
71b687a44102e2311ad8c122cc505b454acd944c
83988 F20110330_AABRYH gezan_s_Page_022.jpg
5f7e945ab13a99bae1f9b3f399ebcb6a
2835565435a8b3cc597a98f6c4486dc9e972274e
4447 F20110330_AABRXS gezan_s_Page_002.jpg
4380e41fadb03b3f82c1dd3d712ebad2
39c9ca30df87089273a66ecb2290925c4ecc5991
5669 F20110330_AABRAL gezan_s_Page_113thm.jpg
1d81a20ee8310c0e8358dec126fc2c13
998c048b935771a980e52a9ca05cf4cc2fdcdfbb
698398 F20110330_AABSEC gezan_s_Page_121.jp2
0a0b167cce690e767a36b5e68f00e713
99f10bf31cc1775896a8dc0af537eff8f2b952cc
114051 F20110330_AABSDO gezan_s_Page_087.jp2
8cd764664c7edf6ed00b25c2461a6184
13881413d7c3d3c2a8114478b8172a303fa8986b
98624 F20110330_AABSCZ gezan_s_Page_051.jp2
6297e0b7bc0eb7fdc4e5245dd946f92d
390fbc38ce1f4b0de81d12d725432e9521946d76
26321 F20110330_AABRYI gezan_s_Page_022.QC.jpg
541d1ff96e03a409df8459bebf9280a1
0fab8624ea7252f75486f8af30b92881a3f3e9ec
22532 F20110330_AABRXT gezan_s_Page_003.jpg
ccdee93f7a0b69e882d1b8b750878aed
b8e564b222eac57b1f7d53d83f1d3f9eb075d3db
50557 F20110330_AABRBA gezan_s_Page_124.jpg
73737251839aaff50ceb9978e6ae2bf6
58c222c2c9c1f0fd088e02d275fef18e55501a2a
22755 F20110330_AABRAM gezan_s_Page_071.QC.jpg
a0389b96055cde6f4f02e4a604676ac4
8323f1e18c8665bdeab6453242423d4bcf45c6cb
38347 F20110330_AABSED gezan_s_Page_125.jp2
ed89083c24fb5bc20d6a38096df52a50
1a1166d5c0ad6c75afe1b72ba48290c62a904327
1051936 F20110330_AABSDP gezan_s_Page_088.jp2
4b57711d5d908a05aad6df7c3b97e77a
db6ab2c7bbcd070c1a704cd07217f1bcb7200aa3
25359 F20110330_AABRYJ gezan_s_Page_023.QC.jpg
ddc429e346f2ea2966011b29f06fa3bc
f0e721ba71a076fb861a76a0344fde7b97a771b7
5965 F20110330_AABRXU gezan_s_Page_003.QC.jpg
6ae3e467a207bc2091c62fb7207d9849
0d10d7e3b13357caffaf8798bea068189c38ef21
8632 F20110330_AABRBB gezan_s_Page_042.pro
2ae31fdbec0b9d7824a4c30e3dcbf046
40b8d053d81a9b73f71f260dec5ee576578a5ead
63841 F20110330_AABRAN gezan_s_Page_122.jp2
817c0b95ca2282ab0e25c33f7c31f461
173275b1d60834db48a79ce08494425e7b6cabd5
92617 F20110330_AABSEE gezan_s_Page_126.jp2
4a2f8d0df3ae5b2ca02a6c32d2d46e7b
153ac0c4b83bbeac86e41a165040a28718d891b5
113265 F20110330_AABSDQ gezan_s_Page_089.jp2
93babc86dd6b2f95317872aa68a92a70
8621f2e1e1a43c7fc6da2f0bfd35a09aa5136567
83182 F20110330_AABRYK gezan_s_Page_024.jpg
ca36aa3bc3853322d07a2e8b332915bf
02f5c573ee626519aa97707aa0b6b8e6f536bd24
19613 F20110330_AABRXV gezan_s_Page_004.QC.jpg
b72150ff935508cac21d99db32c4562b
f9846440bec15c11422f39035090ba071ffd8fc7
58227 F20110330_AABRBC gezan_s_Page_130.pro
4542cfd6a27f8439c669f963209d6c21
9e22c98dd0593a769a6156c8508f26f21fd1b19d
F20110330_AABRAO gezan_s_Page_044.tif
3f721859e4958ac071df073b33b3eab7
53d287c2562b855d4a7d9daf989135462aa0c6aa
128117 F20110330_AABSEF gezan_s_Page_127.jp2
629e12c9d590aff258bc3415137575b5
a2905197219d237d4a56981aafd33d355d918703
861672 F20110330_AABSDR gezan_s_Page_095.jp2
5d252aaae5f846d41a250973d9511ef1
6f267df0d43f1abe88bfda9f27915f63da722f9f
17193 F20110330_AABRYL gezan_s_Page_025.QC.jpg
069760d59a77904d2e2b8176b9665136
45743cabfa18e7d83c9bc730748ca3e8aff7b1fe
78200 F20110330_AABRXW gezan_s_Page_005.jpg
484e6e678e2e0987f1665ebd16ca6189
d1f923e3e8281e277fe726ccba580d197113ce5a
51700 F20110330_AABRBD gezan_s_Page_066.pro
7903f5f20b31c8be0e37b79e94460b06
49b742a7c42ada11fc26c9f4ae74476e07e2eeef
F20110330_AABRAP gezan_s_Page_129.tif
a6447da0e94725c8b47f1ad5bf69692b
bb3eee34a1eab3a0bb90bf07341fe16cf31f639b
119146 F20110330_AABSEG gezan_s_Page_128.jp2
1aecb55191692b37c04d7a5e5f83b715
88e72b33b7a649b4c0871632db2bb313782726cc
105028 F20110330_AABSDS gezan_s_Page_096.jp2
49da5814eff6af13f3a4b313ce05b932
20be836b06fae7c351904e7f42608b11b225d27a
59987 F20110330_AABRYM gezan_s_Page_026.jpg
832ecceb68aa6c3a75a9f9536b3769ae
37eeeacd93bcbaff816732ce74045acab91950f7
130199 F20110330_AABRXX gezan_s_Page_008.jpg
17e962d69f3a49a1f697d4d782afeb39
58ef0c0b0ee5d2096447714c06ee654d81c0b705
58980 F20110330_AABRBE gezan_s_Page_056.jpg
1cef889146791d30552730daec5693ab
4bc40d1db7d1a911aeb75233d42917174e5303ef
F20110330_AABRAQ gezan_s_Page_113.tif
115335c8fbac249e79fa23e5ecc5cc50
4d25c4d00b9398834e4de25d42e3cc06f466167e
22143 F20110330_AABRZA gezan_s_Page_047.QC.jpg
f4dd6a1366792696eb1e554211f8031f
dbd2a11340ffd4b75f9034c23f71b94d953620e3
120767 F20110330_AABSEH gezan_s_Page_129.jp2
4b6183113c1fd33983f5461e6bc4bcdf
da36ee41ce98cd4ae957563fe9e47accbc7f0621
88951 F20110330_AABSDT gezan_s_Page_099.jp2
a4190ebfc7fb2d3d7967688ed9663579
122f1fce579587cd580b43eb6beca1d39fadd4e3
26858 F20110330_AABRXY gezan_s_Page_010.QC.jpg
46e433a2134ead35401ba2d7cc0e1068
e3fb4d68e9203c7b5209f55eb619c231b2ee4f63
F20110330_AABRBF gezan_s_Page_098.tif
bf0ff790ea5bd1a1c10d911519c6ade2
1d391c933befa1a6a9de0b9cd5065866a95aeac6
24922 F20110330_AABRAR gezan_s_Page_040.QC.jpg
5dfef08b38ebf4e4db33b87396bb22f3
9166ee39a39aba67869cf48dc118d23d55e72ca1
26737 F20110330_AABRZB gezan_s_Page_050.QC.jpg
d7c6a18f3ecacfdaddb3a925e69a6554
045fa4232efc1e69bd8578af85b87e5560723178
26228 F20110330_AABSEI gezan_s_Page_131.jp2
7752af7816894994b77c9069b9640f97
e7ef94ce872feba261e6f9d96faeef9621ed8a02
111004 F20110330_AABSDU gezan_s_Page_102.jp2
cbaef47d00d5a65b0e6b8f2ca9430cd8
94954b5c4d844d7288dfd5513ebcdfb12d59e158
18768 F20110330_AABRYN gezan_s_Page_026.QC.jpg
9d3e02d1869e979400ac06ce19d0ff18
12891fcd878d56ae57bf44d13c203173bb9a3224
20253 F20110330_AABRXZ gezan_s_Page_013.QC.jpg
7007d6ad5a720a88269182f3c8f68aa8
0cfcac4520e1405e65e289491e3dcdfbf2401814
110491 F20110330_AABRBG gezan_s_Page_092.jp2
09ce75bd8af9efc42f025ee1d1d80dd9
6483f7f22495b3c9e404bb645c2feea84bfc5153
68387 F20110330_AABRAS gezan_s_Page_112.jpg
ed669c49b976126118e6486eeca61b82
1507f33d2ba2a614417c9c65900649acc58efc2e
77216 F20110330_AABRZC gezan_s_Page_051.jpg
0eda04cfcd78d67fa5dafe76d410d14c
b00b4d8bd38d19d36a4dff990e62351655cda562
1786 F20110330_AABSEJ gezan_s_Page_001thm.jpg
44ce8e522931061e2e830658d252ecff
83476e6381501505582172e8009c6e2ea818f0ae
73437 F20110330_AABSDV gezan_s_Page_106.jp2
c06b197cd573c3e18d7dfbfc2aa1c639
c2482172549c7f79b34593b4c626f97f537b1995
70157 F20110330_AABRYO gezan_s_Page_028.jpg
858d33f9897315409dc8db59535169e4
b4709bf1a6636176e2af35c76560853b61cb5081
6296 F20110330_AABRBH gezan_s_Page_108thm.jpg
d9287a9aea6f4c6a4bab74cf972c48f6
7e330db9b1f294c3ec3ab3d9a058051c9363b019
2572 F20110330_AABRAT gezan_s_Page_082.txt
7d1086c3747d2498e9c7ab35f04a8d4b
c4a3dc601206ea82ce2fcd15f235ec6edce42d1f
23566 F20110330_AABRZD gezan_s_Page_051.QC.jpg
27d88e1268eaf38ca70a6d33e2f1d906
a9d1f702207d7998287f0ae10587ae2c97e3a390
551 F20110330_AABSEK gezan_s_Page_002thm.jpg
2b01da536c567b1a7e62829d4ec244f0
8a30796badc2afa63a8938438a2bbf41672377ac
104589 F20110330_AABSDW gezan_s_Page_108.jp2
cbbe5c08902200352c8de06490bdd111
7718bed95306d4c9d23f26117cdfdf48b58a9752
25663 F20110330_AABRYP gezan_s_Page_029.QC.jpg
410fc9294ce127c8f128350de5f6875f
50f84509f4a0e352d363542d215dda80db4f31db
F20110330_AABRBI gezan_s_Page_002.tif
82fca7ec9796b38a901eb8317f2db78a
bbb8a095a2a2634aed8cc70c3dcd536e1e0dd9cc
24559 F20110330_AABRAU gezan_s_Page_105.QC.jpg
86d519d976a2d123fba3d67bde3ffbf3
12e916ecffadc6c5ace149604715d31df9339481
38272 F20110330_AABRZE gezan_s_Page_053.jpg
e1c4afb46be9b6a89e319e3db3982014
267e91593e3c6f9e279cd77e083fb41c10b6484b
1636 F20110330_AABSEL gezan_s_Page_003thm.jpg
73a89811ad22abadde8c2945de19d2a9
bde4cb1211b953f88af7f2eb9a92eeb32db091be
81673 F20110330_AABRYQ gezan_s_Page_030.jpg
6e4528423847c6a219ddaa5c211da81a
672e53c106e5381bb7598c9ca3c196f9cc7d34d1
4217 F20110330_AABRBJ gezan_s_Page_053thm.jpg
b1080dda0c1480c0295916e2a1f659f7
85c06abe02370746a58d83244e352acb9e23b3c9
13453 F20110330_AABRZF gezan_s_Page_053.QC.jpg
815b626fcf9d5902163d13e814d73fd5
dbf835143c989b9e6dc54542c6c8359654e85b26
5874 F20110330_AABSFA gezan_s_Page_038thm.jpg
c2c38c8318c6993b18a27d9e78c768ca
601e4cdbc1fc74609ff188e64b641e133aa72bbf
6721 F20110330_AABSEM gezan_s_Page_008thm.jpg
7205ef30b337575530e3849ef96eacc1
06e93311546a46eb40f8c535845a8f3fd0a0690d
105295 F20110330_AABSDX gezan_s_Page_109.jp2
7d425fa054420ac83429644be4a05e70
9074321418671ceacc28bf423b6636a62c2fb413
78482 F20110330_AABRYR gezan_s_Page_034.jpg
9e2c80f344547c9d26932404653fe3dd
5dbe8688aa6cab86de4e00541af1d6a2f63aa392
1973 F20110330_AABRBK gezan_s_Page_093.txt
77f9d075810d02d990c0f0c5ff526483
65826fea91df439069364daeb0f58a3446e2360e
22583 F20110330_AABRAV gezan_s_Page_043.QC.jpg
65f6e7daf264eca438775dc5520db785
b83f9a44a232597cb45ab20a2caf0d48cbcc0872
77353 F20110330_AABRZG gezan_s_Page_054.jpg
e8bbc97174e475aff2ac9d78dde984b0
7853d7b9a6eee49d7feaefd55abccfef47f7c1ee
6371 F20110330_AABSFB gezan_s_Page_040thm.jpg
c08a9fbee422cd88ecffe4a9b1bcf809
a775c40f48b3619c9a9fd81ae930c80d53242e25
6259 F20110330_AABSEN gezan_s_Page_009thm.jpg
e133c738fcf05323232d836191c59fb7
f15ba1c118dd13beb0fa2cc2b99bdb1704a339a5
103719 F20110330_AABSDY gezan_s_Page_116.jp2
412da1fae34e243872c76e0ffc1dcd58
a416dbea0deff253c67f12d4c915599ae21b00f7
92034 F20110330_AABRYS gezan_s_Page_035.jpg
9eeab156bde0409851fa20f040943851
01166551aa4cb43078c42a1b1be83580e56285b4
F20110330_AABRBL gezan_s_Page_068.txt
e0ab94ce1291321e99238966ff10b1d4
90e76ed6e9bba554575420faa77e1c7d29a3f70a
5578 F20110330_AABRAW gezan_s_Page_016thm.jpg
345cb2c026be75751a827bff3f645271
3c602d34a8f436de3126204d96894749e4ef08fc
77349 F20110330_AABRZH gezan_s_Page_058.jpg
cbf2c2d872aac1a2e1bc63b4a92c3336
db2916e537038dce9685013d71b926ce4ea40b6c
5838 F20110330_AABSFC gezan_s_Page_041thm.jpg
6683eab77cdbccd0bb78c449c32e98e6
4a8b2e89ccb12cd84393db9528308f51dae3806a
2569 F20110330_AABSEO gezan_s_Page_012thm.jpg
901432b584f57b01be08d2350890f62e
fc1dba51610d818dbbcc232196320bc77670748e
72560 F20110330_AABSDZ gezan_s_Page_118.jp2
2cff9771c7138e00450d80b26408126e
6261bf9af1ce1ec333720cb5c2592b985e49e3f5
77955 F20110330_AABRYT gezan_s_Page_038.jpg
46ae91ca8e0f591b39c1d3a64f7de659
3a73333aa7d47d71f094d975baed8400e683ccba
45910 F20110330_AABRCA gezan_s_Page_061.pro
a2d4b5a2de6cb6cf65c8cf32604d535f
3ccbcd2cd09afbdd6a4fee53e18fbdbf2739331e
F20110330_AABRBM gezan_s_Page_132.tif
09f61ad87bf601b00c037d898bb07946
a9722b861bbb7f1ed46f40bd9ade4b5044819341
2882 F20110330_AABRAX gezan_s_Page_051.txt
2a87c4e3443532b5eb5a87f9ec3396c0
00ee9816d2dad2b05fe98fa616026dd0958b4073
54454 F20110330_AABRZI gezan_s_Page_059.jpg
7eca1e9285811e209d44b74e6dbb76b6
809fef1c322f59bf1fe64ecde4f38ce6fc06fe79
1498 F20110330_AABSFD gezan_s_Page_042thm.jpg
5db551782d0c6f8234e5a6d36fc30b57
9c6bbdec8b619d89d57254f36368808a5ef1c7ad
4796 F20110330_AABSEP gezan_s_Page_013thm.jpg
cb0e4fc2e3ed88e522b803421eda6739
78f3cd4818376d8c836f32f0ce6b464c7eb4fba0
24686 F20110330_AABRYU gezan_s_Page_038.QC.jpg
e789a05adccbded953bee1a2bad45ed7
fbe01b0f52cd547eb6352d3ba4f8eb8867896d6e
1918 F20110330_AABRCB gezan_s_Page_073.txt
9c9f8fc9bf1de3711a149750549aef9a
c676edf4695e4f335aa16d955713e147698e06ca
32892 F20110330_AABRBN gezan_s_Page_106.pro
a5b7ecb56ab436735524c7b8349bac33
6b58e9f64cb9376a388c0047ad637166ddb706ed
79719 F20110330_AABRAY gezan_s_Page_033.jpg
b2dcf2b4656ede9b721f7120499876bd
7ef9f10597270f4412ab6fa0b89eda560da454cd
18491 F20110330_AABRZJ gezan_s_Page_059.QC.jpg
bf569a3e5fac8da886c09cb008da0181
46795baf99f49eef56daf2b6e11e292b08d7084e
6063 F20110330_AABSFE gezan_s_Page_044thm.jpg
7590b005c7ee719331a368447e3cf031
9ea72bf61e96e57030e686b7ffd039378f3d9c68
5670 F20110330_AABSEQ gezan_s_Page_015thm.jpg
21a5d0c9b64f0341a26a4a696381ddef
3cce4f495a0d76f711138e4573bbc58e21c0d262
61440 F20110330_AABRYV gezan_s_Page_039.jpg
860df59fd6e29da77159c736521580e1
2c6cf7549eaaf450ad25f38148689c6a6e5c51e1
72530 F20110330_AABRCC gezan_s_Page_043.jpg
2ee550f354fd721eb874d60c8c9d37c8
b7dec63fd489485f87171e3199efdc243be4d9e0
1843 F20110330_AABRBO gezan_s_Page_038.txt
b7cf823b4c582b43b04bc32bb45b4a01
9559eff2ae740cee6ecfdf2db03e1cd2c987ba30
22138 F20110330_AABRAZ gezan_s_Page_101.QC.jpg
5fe1fef40a230292af9cc6268e507cc0
e80b5ac6a422275727276d61e9095b74a390ae2b
62395 F20110330_AABRZK gezan_s_Page_060.jpg
19df51486a4fa54198f0d563a4b73fb1
23ab2f7aea46ec292735bd5dc802eb48ea9c125d
6213 F20110330_AABSFF gezan_s_Page_046thm.jpg
d31e423ab2e10380b6845286a71765ef
9d06955ca34199746b2a3e938a4e2b7d7be9ccac
6140 F20110330_AABSER gezan_s_Page_018thm.jpg
580b0da0d7fa6018264850d8cadd7f19
1a0d476a2074621aea980ae9beb1d52d469cf224
83420 F20110330_AABRYW gezan_s_Page_040.jpg
161f6e7630d52c15ca1d810c8563daa8
4465c8f5089df29c19152897894aa9f3290cab6a
39318 F20110330_AABRCD gezan_s_Page_059.pro
067c37fc39b6ae5bb33862674dd3ea56
ba9a97440c3bed88327a24ca78de916e37abc2fb
21842 F20110330_AABRBP gezan_s_Page_115.QC.jpg
36b93017eab5ec59c9ed9be51fb5a3d1
50b4e88159c639a0be2ac7901182ffd740c92da4
25082 F20110330_AABRZL gezan_s_Page_061.QC.jpg
67923e6611a9aabd1676c0fb159d45a9
f102790ca0f900b8e427b9914628b233d6cfbff0
6168 F20110330_AABSFG gezan_s_Page_048thm.jpg
06eb956a2b790b1802e52dc13e415e79
d88a8b3e4eb7896ef66605c9f1f613d89c758d4d
3062 F20110330_AABSES gezan_s_Page_020thm.jpg
3444f84b05d455f38a722704ef711eb4
50b9776f8d557f4f945030286b9aabc0e02a1cad
24349 F20110330_AABRYX gezan_s_Page_041.QC.jpg
5a7ecda855a62e94a1822b53ee4cf4e0
eaa9d1d01e9fed3c0f667db00f5494cb0ddf67e3
F20110330_AABRCE gezan_s_Page_100.tif
f6b4d522f465a080a7fe6ad8ca1c6e29
ad7f51123e6b17e481d1b7abb0c9e436f60f9798
25005 F20110330_AABRBQ gezan_s_Page_007.QC.jpg
15a6bcee4097becca43f5411df6928f3
3dfc919a8fce76422811a3c80a2614aa757d5c71
73217 F20110330_AABRZM gezan_s_Page_063.jpg
90326e9ca7873ded154650441edecf0f
e88234884c32aee4373b89b759355fc18753d877
6488 F20110330_AABSFH gezan_s_Page_055thm.jpg
a8b420a65ec0843d9c1d1cedfb876133
74880b8cb04c5cfd7bc036ebaaf54693f6ea456a
5618 F20110330_AABSET gezan_s_Page_028thm.jpg
46e200eee890415fd7d89e53eda224e5
cbe04353ee037d4d09c1ebd20a159a8f970fe8dd
16838 F20110330_AABRYY gezan_s_Page_042.jpg
15a5df8742f188c3e1454ac0f83a0431
127f4e9ef5391bab6120ee83f03d9723e7bf07f0
937 F20110330_AABRCF gezan_s_Page_052.txt
221114df76ee3f83c6381223e316edfd
713d17d8b30d1052705f5c9c56747d563bcb6c5b
62575 F20110330_AABRBR gezan_s_Page_118.jpg
1f866677e6a90eb61603a3150e7a1a31
e53024f98dfdd103ba6faab9e80f085673aa3d69
26898 F20110330_AABRZN gezan_s_Page_064.QC.jpg
055bd92ce61960f265cdfb01638b6082
77338752b859817ea34259002842f74be0a52d87
4817 F20110330_AABSFI gezan_s_Page_056thm.jpg
9fd1dc90dff42cd9834f363e9f891ac8
b4838b5d3fe10698f5adc277b239159e22f54999
6204 F20110330_AABSEU gezan_s_Page_030thm.jpg
ee81a250e37a8bda92d1d1cfad37982f
9dbf3059fcb99d7dadb64c7d3171857eb937058c
F20110330_AABRYZ gezan_s_Page_045.QC.jpg
55d6f68e6106222f71a1341683d82ca1
8cb28413e5b4776b53a26a9ee63885ccc0c68a05
22227 F20110330_AABRCG gezan_s_Page_063.QC.jpg
5bb37c8fbf3e748613db625c684ecf5b
72af990deefd36d02946d2f77415bd4596a72f7c
F20110330_AABQXA gezan_s_Page_024.QC.jpg
4062eab892b5b62772e8776789ff469c
42c9e1fd50f2fbdbafa897aae5bc559f62ef2e72
6137 F20110330_AABRBS gezan_s_Page_029thm.jpg
8eb9c9e9cb363b6f5e29398e1a66e9ba
8b5245d2631de79ac462f16e1a6ead6b85fc0072
6208 F20110330_AABSFJ gezan_s_Page_057thm.jpg
8ad069c4d252be895c8e1785732b4678
771bcb9b462b0c34fee3bbcd0ee1c578b35054ff
5553 F20110330_AABSEV gezan_s_Page_031thm.jpg
44c68d20b6ebebe7a91d4349aa67a500
7460facaa2e37edb221d06929aeab81ebcf7100d
F20110330_AABRCH gezan_s_Page_062.tif
20c2bf5a6c780fe2f82f3466207f83df
835a62db6f95ef66b43e22d96e83a2e73c9e10d2
81180 F20110330_AABQXB gezan_s_Page_093.jpg
6fcbc9acb5a9e620059e700484dd8bd1
0d9a8fd79192342dfe8f32ce792f2a8466f490bd
F20110330_AABRBT gezan_s_Page_077.tif
e19e15bc4433c0c8b9bb963c00b89264
4d7fd16f9fd99560c9582cd8e86b5b80858eae98
24787 F20110330_AABRZO gezan_s_Page_065.QC.jpg
082b0da83633119c497b1e81cae41f66
39738bb8b648a63da6adedb9d8290f0b4c8b5597
4930 F20110330_AABSFK gezan_s_Page_059thm.jpg
ab6a77a573f9ac4d2c3b374f9b70fa4b
1858ba6d9d280078af847fed84b6cfc8c7766d4f
4396 F20110330_AABSEW gezan_s_Page_032thm.jpg
e75a13f269f0cdd0f7c40da1df189b76
7166c8418da7e86dee433c3fe58faa53190583bb
24328 F20110330_AABRCI gezan_s_Page_054.QC.jpg
7d3b255e2eb45f46aa83e5e333ba883b
11756be4841b015a4565618e1eb48304171a0f37
49075 F20110330_AABQXC gezan_s_Page_065.pro
75847a5f66558fd582f0a30d26dea50c
80ea2e43478fa71c9c277a52c4febe4c96dfe388
F20110330_AABRBU gezan_s_Page_048.txt
f578145789adb46614f15a1eceaf8885
bcaaa41ceb7555f8910eb8bb1074de909d20e516
84179 F20110330_AABRZP gezan_s_Page_066.jpg
6bcc32eb11a1e758b8e903e68fe56bbe
c015a8c223afb028f889bd3cb5f16111313d3fe4
5056 F20110330_AABSFL gezan_s_Page_060thm.jpg
81b26932612cdad1739ac2b1c76b2b93
de0981b56fd539a8a2a23344c5d52f3d3c789eea
6174 F20110330_AABSEX gezan_s_Page_033thm.jpg
71dc2e034f81d8e211ab29c0beb5f6b3
d6989461f6c71aba7809e5222d8d97a76a307690
42836 F20110330_AABRCJ gezan_s_Page_063.pro
ce0673140d173446ca8837a01459baac
0829c18651acbb1d4dac6598c8ecabeec52f80c4
28088 F20110330_AABQXD gezan_s_Page_009.QC.jpg
1893e8174cf3f695dd188a935aa535a6
1fc715192ac3c62c0a09e34e0c5e08807ace83e1
24259 F20110330_AABRBV gezan_s_Page_074.QC.jpg
8e8c87a01852ff7fad533feb263b0930
05d162dbbb51cc2883f26c282aa0d2b7aa8827cd
23956 F20110330_AABRZQ gezan_s_Page_067.QC.jpg
a2fb766c8c3d5a47afc7ce1ee2f51527
477dbb0bc1ed43b1470ccca0dd7fb0a34d73a76a
5087 F20110330_AABSGA gezan_s_Page_095thm.jpg
c63eb6531becd40797df9e3aea0285c7
2e180f8fc00d1678c94a831db71ea88aa0f9b888
4997 F20110330_AABSFM gezan_s_Page_062thm.jpg
6f2cdc5c2313906055f8bb938a03158c
bc46fd21e563758c7bad9583866e566d9aa2c2a4
51161 F20110330_AABRCK gezan_s_Page_022.pro
1a776b2c3a6d548921ebc24e1c09ce15
95c1e4695fd91383883a54284217455b906ed73d
38235 F20110330_AABQXE gezan_s_Page_032.pro
f8821bddf0abd1c6a10b6470b7f114c2
fbb146d829980b84c4f3047f1e31bc48f91797b1
78351 F20110330_AABRZR gezan_s_Page_068.jpg
1d5226ad0ff01f4c2606044a81ff8a65
260f2d7b0efdbf5314cd44f15e7b6de39fd6611e
6026 F20110330_AABSGB gezan_s_Page_097thm.jpg
b75385e8b5bb7c611d4580be1d07c9a8
74799d24cdae19bfe85712c1f26408d65877fc6a
5438 F20110330_AABSFN gezan_s_Page_063thm.jpg
7ac070acdc620e649274cc13cc400fba
9ab14fa05c51b426169469ed4e980692f4a6d668
6545 F20110330_AABSEY gezan_s_Page_036thm.jpg
5517fc253075f39622436700a918a4ba
6a17f9624432c4784c76d53050fdb3f041c79636
3034 F20110330_AABRCL gezan_s_Page_113.txt
3c7e8ebf246175b31bf3f6ef1e2de85d
cc6f1089f5927f79250b224d7adb3a54d9b51f79
71570 F20110330_AABQXF gezan_s_Page_101.jpg
6667e57c710e8d591ad0aac631b3212d
b08a7bb2d0c5c8c220d91215f168a0406e2c5d50
22912 F20110330_AABRBW gezan_s_Page_028.QC.jpg
1e2cb35c2e5a62c93c46ce20672d7dae
342fa43aa517d26093f05e9ca1ac9ce45f9b5616
24434 F20110330_AABRZS gezan_s_Page_068.QC.jpg
91e3a398fcc407dd45be9a8d9702477a
4246c2b1969d7892df7193de4d2d958b8284d3a7
5826 F20110330_AABSGC gezan_s_Page_101thm.jpg
91eec02decd185d7a682a7d46ccb0de8
bc0a2f2f94e424c8362863804e8cf5d7e0b3756b
5786 F20110330_AABSFO gezan_s_Page_067thm.jpg
6f54a38af7c5d157ee5a0cfb87ea6911
07fb2c7d130732445516f0ea0c80fb0364d5d921
6031 F20110330_AABSEZ gezan_s_Page_037thm.jpg
3fa9a99244513ab624175d725405764f
61219a12cb33f4bb82c8d50e36e5ee8e4ebaef4b
58370 F20110330_AABRCM gezan_s_Page_076.pro
a32ed288e710b9bc44e9ae369296e34b
931c0603bbe505ac74ed11bfcbb8c1bc61656245
106516 F20110330_AABQXG gezan_s_Page_085.jp2
f189e3b75c35827e90d15b854c2fcf9c
8743fab44e8e528ef691babc245cdfe656dfbe56
1604 F20110330_AABRBX gezan_s_Page_131thm.jpg
dc08a0c1fa6284f201613f3d15cda643
d775651358133f194405f5602e6180b7c8f8f2f9
77979 F20110330_AABRZT gezan_s_Page_073.jpg
6df32e5f37eb6f35e771e2fe03f1f00c
8f273eff6a834bbf80c49c70f99bbeea374ec2b2
1859 F20110330_AABRDA gezan_s_Page_041.txt
14f489788f4f9c0d7f6e04e60ba3c5d2
48bb310b0f4ecbeb551d2bdb80fe49b87bde7491
6421 F20110330_AABSGD gezan_s_Page_104thm.jpg
256f5ab20fbf25a7f0a0535b915fdb8d
1fd28b29cc23a0265391f44e53050ab87bc292c7
6056 F20110330_AABSFP gezan_s_Page_068thm.jpg
47585ef0e2d3d78d0af2f194c72bfb97
1aff87a60d43d6e20dce71c02a9a152416355023
5519 F20110330_AABRCN gezan_s_Page_126thm.jpg
d4682dfbc9361a907d2991cb9d945e35
ba5b50ed59a25f948fd913b885ab366aec648cb4
519547 F20110330_AABQXH gezan_s_Page_072.jp2
ccf374ba40923423e32eda2884870ef7
601ca31f3aa495b34e850e3eb214893ffbbdb763
6030 F20110330_AABRBY gezan_s_Page_054thm.jpg
0a4f41362636964a7f1309bfc8ea40d4
2f0764ac458bc1dec196b7f96adba6b25c490711
79592 F20110330_AABRZU gezan_s_Page_074.jpg
05654e74325f77c0ab37bdb794208ea8
be2b892dc5308ce0bfd43ff52d072fa26af5153e
F20110330_AABRDB gezan_s_Page_114.tif
24a9907e956b702586d9202a7c4bac53
5a5fa6650c3fb54fb96346340369ead1fce90963
5920 F20110330_AABSGE gezan_s_Page_105thm.jpg
d0d7ce638923a5ff24b61ac117905284
d4aa92aeff00bb9e2b2801e60ae404b4e1430cd7
6145 F20110330_AABSFQ gezan_s_Page_069thm.jpg
31b1dc88cf98db95e0354e26a476810f
a9d430489ecea25a52e0969ca64282dd8169a15e
F20110330_AABRCO gezan_s_Page_080.tif
3efc53ae331d4aa37d4c85da28d1daec
3976785bf6a279102c37082c902c10bfb61981cb
53412 F20110330_AABQXI gezan_s_Page_114.pro
e327308934a2009cd0ce57e02e1f0e30
2c3276e039ac758989ee6fe56e8720711f213f54
3075 F20110330_AABRBZ gezan_s_Page_124thm.jpg
53c9c7f34ec90e95e0e1be30c3d06049
b3d8cfffdafde16ef9f0398e368f0c11787e07df
22558 F20110330_AABRZV gezan_s_Page_076.QC.jpg
d80ab17f6e4dfbacd44d48fb421435f4
1b165d989e2e7763e76a02a492c5971ce105e4b4
26383 F20110330_AABRDC gezan_s_Page_049.QC.jpg
1514acb517c7451372eec0eb4fd05955
2c5c2f37dc6fef3610bce27163f3781c5002b6fd
5241 F20110330_AABSGF gezan_s_Page_107thm.jpg
44eb044e2f267a2e2a3215f11fb44973
a012ede945ba6b998044a7f9a12ecdc9d4bc6a2e
5880 F20110330_AABSFR gezan_s_Page_070thm.jpg
688eff2f0c16a4889cf2ef9e14aa920b
187a2cc1fc9a1f76055164b8a79d139fb00d9538
55478 F20110330_AABRCP gezan_s_Page_106.jpg
5495ef2a7a29bec45dd651d40e6b66d5
c2695d5f12c62193fad84d46646abf67d7b96565
50886 F20110330_AABQXJ gezan_s_Page_036.pro
a6f801e52e9b7479b028752ba3a9b59a
2a3dea7d4da29fc808490977a71c0026bd5f243c
23199 F20110330_AABRZW gezan_s_Page_077.QC.jpg
e92bc9d809389133fecae57ce2f80894
d386cd61a0fd155c3e4d81e8de87c22dbbe30464
113587 F20110330_AABRDD gezan_s_Page_064.jp2
329a117e5d6fa1341b0e81ee70b40cb3
93ddb501db97440ea37f7146ab01507dee64271a
5267 F20110330_AABSGG gezan_s_Page_114thm.jpg
9c2dc8109ee23d24d669ef629b475b1e
85de3650c5a291bd6924bf27fadc71a7c080696e
F20110330_AABSFS gezan_s_Page_075thm.jpg
d8e623f9bcef91d5c7a4c104443e732b
7c54662faa756368b67cd11e53638dc73c789e8e
1819 F20110330_AABRCQ gezan_s_Page_080.txt
9f85338138a8a77b56cab09721b34134
b6daa420c608f869940370cdb2ecd5a5abd6d846
2548 F20110330_AABQXK gezan_s_Page_127.txt
c95b5ce221118254bcb7231af0109c4a
0176c8917f9f3f2813924ed5c6b2ee3fd01bf4ed
77790 F20110330_AABRZX gezan_s_Page_078.jpg
39e64a335836db172be2c16416142bd1
7bf6558143a05c11c7a655a17af6f9ca35116bbd
F20110330_AABRDE gezan_s_Page_081.tif
57edbeefac183f7139c885388e9a45b1
5d574f48ca4717d5f2db031986a7650a1d1d9d95
6081 F20110330_AABSGH gezan_s_Page_117thm.jpg
4cbf8a4ca8659b1e6734492459ec9afd
0a0a3be66264e6574699c62bd65afc48fe7b1406
5536 F20110330_AABSFT gezan_s_Page_076thm.jpg
b816f27fc2524e243dff7059890a1e73
01c1a2fac9b531fcb332f713b871707f7fbf0d0d
5541 F20110330_AABRCR gezan_s_Page_021thm.jpg
f8af7642fcc689c3d101de295db35a32
959f62e108a042a48b78e8f242a31c994da3bf2b
113892 F20110330_AABQXL gezan_s_Page_017.jp2
e903f7f741b2b94ca895f4f62103de78
c833297aea3dd1a0476cfbd4e240c7f96aa68e7f
23502 F20110330_AABRZY gezan_s_Page_078.QC.jpg
7d56015a2a79984a14b1236816d85445
cffba1abf7e29f37e32f5278c36a92b209883886
18873 F20110330_AABRDF gezan_s_Page_132.QC.jpg
c229aa4d93bc3c0598af8271dded059f
51a49550793b396cfd1b2f5a116df14a84dc470a
5855 F20110330_AABSGI gezan_s_Page_119thm.jpg
bc5cbdecf9b6e901734cf6b3db65ceaa
8a3673a2824cfc3b9c1f030c6632e76c3435051e
5911 F20110330_AABSFU gezan_s_Page_080thm.jpg
73ba90e2c4e1ce5aa54854bb9baad618
e16c20ae9597f4419101caeb2d3193d5deacdb7e
1640 F20110330_AABQYA gezan_s_Page_107.txt
eb2d5db5933e39700a6fdbf6b87a4e68
6a1f19b7d4b4c279219e583f64397641fc3bc1c6
24578 F20110330_AABRCS gezan_s_Page_097.QC.jpg
138424cb379aec09d8c2b993de1cdff5
80b360fed83c1df619f3352672694de48eb3b668
1805 F20110330_AABQXM gezan_s_Page_032.txt
becedb60c8d646c8ef555535bb993545
372da6e6f867899ef6104d9785c6d70cd3617d3a
F20110330_AABQWX gezan_s_Page_009.tif
7119263f6c1645d0f426968d5b2c3f6e
df081a83189a319d2ab7479e2e5f97316bfac683
80427 F20110330_AABRZZ gezan_s_Page_079.jpg
b37b7d11b44de06109d8464f51ee33f0
07f61baddd606bcd01d4c667b489533f276ce610
70837 F20110330_AABRDG gezan_s_Page_031.jpg
918504be3fdbe04352373d63c6c424d3
6204bfb8b88260005da8ea20f7a6ab1d79168e25
4210 F20110330_AABSGJ gezan_s_Page_120thm.jpg
845ea0a62cdbfa88393da008496a1be6
5a95576ab87cec1631d95307764a941cbdd00e00
6254 F20110330_AABSFV gezan_s_Page_083thm.jpg
41f2eda2d2342f9bba6a989203d9e45e
3730aab298877a7d874348582fab065841e1e4ef
79692 F20110330_AABQYB gezan_s_Page_045.jpg
38efbee3f5c846421c927d229be004c8
5b63f654884963722ae4b0de77aa9be2806e5bd3
8121 F20110330_AABRCT gezan_s_Page_001.pro
069ca9f3908f3c7caeacfdee02bcf6e5
584be7215cb1b16a656e3b13699b1eeb28f70f1d
47405 F20110330_AABQWY gezan_s_Page_073.pro
a496a8f60ee22418fa43fd6eaeda4936
1ae8b2297faeb384de695e0f341edd049d1ef927
85526 F20110330_AABRDH gezan_s_Page_055.jpg
fee663287f3cd81212307df7fdac0841
7067d9a1c8195afdb118dc46d0be157ffcd14721
2163 F20110330_AABSGK gezan_s_Page_125thm.jpg
a89d653c909263ffe1cbb589163be80a
34b3fc9714458659c8be12cfd6535716961d3d77
6177 F20110330_AABSFW gezan_s_Page_089thm.jpg
c5f8b7261a038dd6bb03d04d4a21479d
ea7b9054f28a1cd9040e58a30a2141bd905f10c0
4151 F20110330_AABQYC gezan_s_Page_111thm.jpg
a25e0d1ea4dfd7ac55f61b8ef40565e9
11a322d0266fde3ab8dd32836809973f13194ea5
F20110330_AABRCU gezan_s_Page_062.txt
00cbabe18a757fac849e8815de7519e8
873a4b85fad7f52eeebd184ec8f3d5657e4f5e9c
91981 F20110330_AABQXN gezan_s_Page_123.jp2
e74e6cc4ec8aa3535c9141c747e64f83
3b3f3033871411c75a200583c95a36537b6fa4df
F20110330_AABQWZ gezan_s_Page_123.tif
f30b81bc08a31945e1b18ef96bd6dd0f
67fe4ed235a86836d5cd25a97de62a326e97ac5a
1906 F20110330_AABRDI gezan_s_Page_045.txt
c8ed00436f96421f901831b318859889
e18c8fd88b4a42c890a5e7c11f0260be97f05e1e
6637 F20110330_AABSGL gezan_s_Page_129thm.jpg
c5cf01341f0519a2f899519d7bdb14b9
00a3dde14872ec6629879309a9b89083a273eeb2
5195 F20110330_AABSFX gezan_s_Page_091thm.jpg
f88215b9c2993ac4e6b717939b7602ff
02428a33a1d264bc38c79806da366e9a7c9fa9ac
102890 F20110330_AABQYD gezan_s_Page_034.jp2
452180f069006cd0267167ef895f96e0
ecbf2d95326371a983ba5c4e40cb488d6d690c2f
6239 F20110330_AABRCV gezan_s_Page_079thm.jpg
69c9e666c91f36a7520a5672d3dd8537
e42349742e9cbef026983bdb1e75169f3a2b0255
F20110330_AABQXO gezan_s_Page_120.tif
9b87698d8ec76a4c590e26513818d3e2
75979b9bfad824a1bcdcb7a4481edeae4283dd3a
F20110330_AABRDJ gezan_s_Page_075.tif
79538f0f4bd85c0a0b341decd7e47f2b
839554d8e82aceedc3c031dc10735d36a82ba6e6
6443 F20110330_AABSGM gezan_s_Page_130thm.jpg
b49caa2fcf117a4ac0102a41ffb5c0ef
ecae647245a21b749e7e5f73bfc7c62cf1647005
6105 F20110330_AABSFY gezan_s_Page_093thm.jpg
71a5b100e6d1c3179c7b920c10936495
90cd86848cd948761b43615320311f196960a7e7
46729 F20110330_AABQYE gezan_s_Page_074.pro
98e05555d44110d34456d730d15c1680
48ca898919eba7c793665380258913e3095fd397
109800 F20110330_AABRCW gezan_s_Page_066.jp2
fd8f265d210c9b7c9b78f19fed1bbd88
ec42ea7280998e86e95185913d7a7ada8144dcef
109525 F20110330_AABQXP gezan_s_Page_024.jp2
6c53aefd646b7800b8932ff9b47949a0
85ff2d1c6829d46e6fa8284278ddf8b640ba7fd8
47997 F20110330_AABRDK gezan_s_Page_108.pro
eadb4eee5e36c9089a2916f4258b024c
8c67d67b4b9e101049b70732aa634bc5035805f0
4644 F20110330_AABSGN gezan_s_Page_132thm.jpg
77329c21eb496438ae81400b5cb9cb14
cf00174d3b9efdfce8e7a8524a07364d21e6a249
83223 F20110330_AABQXQ gezan_s_Page_049.jpg
18171ad75010b0cd571298445292cb26
060c770d54d5ee4e67c852f89caacc6e09b3c3b0
18719 F20110330_AABRDL gezan_s_Page_039.QC.jpg
46ce06978a733ea50827867c3e64dcd6
4ed7a3d6840d9aaad111aea7ce7d6244ccfa2716
53556 F20110330_AABQYF gezan_s_Page_123.pro
00367b8ceb20f9ddc9d2d2e7e197ac06
c1f3d40812c596692d5687da7935d8d60c4901a3
771360 F20110330_AABSGO gezan_s.pdf
28bcc708d71ad7e93b9df66babb3dd7b
2ae782eee3a1a5e75e9330fff3b277085e25fdfe
5714 F20110330_AABSFZ gezan_s_Page_094thm.jpg
ece7e802ba139f2b021a08c2e8b3a981
ada785135e4ebc6524af2ba684e3c91f79046358
27316 F20110330_AABRCX gezan_s_Page_129.QC.jpg
b9d9b8f7e9ab86124355be5dfbd7cdc5
c983677109e3d333a412d8f0322dfebdd17ab14f
25227 F20110330_AABQXR gezan_s_Page_090.QC.jpg
354d5d733515340d0e3b097339a96b5f
b7b06b6c4b0ac574b5775bf510f902518d05d350
6321 F20110330_AABREA gezan_s_Page_102thm.jpg
8481a7a416fe84960f62244a95de339c
c53264e256e7c62fde177023d3798f27ed959430
19650 F20110330_AABRDM gezan_s_Page_006.QC.jpg
c954bdbb3d1d098f1a02c5ca067bdfac
77b0987d7b960f1b4d2a577e80622452f015dadb
754153 F20110330_AABQYG gezan_s_Page_098.jp2
dcf9ad82bc610e24540d78b987c41acc
07b00118a7efcccdadd7a199ca9a3840164549d4
154356 F20110330_AABSGP UFE0012929_00001.mets
b177cbd32d4ea7146b65e65b17d16c31
3918e11abba5e1ee9008788e101af1b500186237
21184 F20110330_AABRCY gezan_s_Page_131.jpg
1964e022ffc778869c7d07a4bcddd104
deeb2280a703154035261ce97d161aff9a5f5dbd
55878 F20110330_AABQXS gezan_s_Page_120.jpg
8b1314b7dc168442728c078404f4fcc2
05331b949aafbf33ed2ff63bd83b04a291fb10fd
F20110330_AABREB gezan_s_Page_087.tif
5186be52a88ea253570f4f25dbd9eba7
d2ff90f3c20b7893a19aa6f336060b89e6ee7768
2066 F20110330_AABRDN gezan_s_Page_017.txt
4763ccf1a78cfb1414e01ef66f714444
ee4427a5a24bcc39d75b6d7a1db13006e620e16a
23137 F20110330_AABQYH gezan_s_Page_031.QC.jpg
ec6519972d870a8646b82237245b1c0d
399e03deeecae54009efb6edfa8c9e4109c8bd55
24985 F20110330_AABRCZ gezan_s_Page_044.QC.jpg
099505409c22ce9710471e8505d9374f
2dce6234dd35cc073a66ceab2a83a488ffde7ac6
924048 F20110330_AABQXT gezan_s_Page_037.jp2
0730472b434fe5999bbce0df2f883c20
50f57266db9c2c80d63dbb07b9a181c716bb3094
1490 F20110330_AABREC gezan_s_Page_095.txt
e768e90bda06502499a9c93b051541da
9947a5d1de074bc61517ad477f8c3f803dffe4d9
4773 F20110330_AABRDO gezan_s_Page_014thm.jpg
cce7caa88413134920597d64a3a00b87
627c40f881544986666a304e4dcf110a44c77c3e
84810 F20110330_AABQYI gezan_s_Page_113.jp2
f1f17852115af4b1d065b53d69383a9c
803f633521723027cdfc79adcafc386fd5d434cc
79535 F20110330_AABQXU gezan_s_Page_109.jpg
12445059d98763b9926a92fbf5f5a422
4893bda8d80d347d36211414af97e09110c08f6e
24479 F20110330_AABRED gezan_s_Page_001.jp2
1443136e1edfdf02666dd6e858dac0c9
5d60af965039a04a831b346c70bc5e8bb8f9cb78
45067 F20110330_AABRDP gezan_s_Page_071.pro
8c0a2f70c88f1c8597721b0d504300f9
47b2b4903168c2f5fd434e2ece51684636d3d76b
1899 F20110330_AABQYJ gezan_s_Page_108.txt
4dd450967a7b49bf7e33393e83d28bb0
901a93cb0d64927c515249e62a42256db096e5e1
F20110330_AABQXV gezan_s_Page_088.txt
99668d9d774dd04fce3fd04b1e007605
c21e063103a977cbb2d73fb3ba919b5f769793b6
1540 F20110330_AABREE gezan_s_Page_002.QC.jpg
ab2d26bb2ac75583c5473033b4c900d5
12dbd50afeaee6304695dbbe291322a7b3d96203
125819 F20110330_AABRDQ gezan_s_Page_011.jpg
508f3dd8858be933721a70464c7b4901
cceae18c76b08cdb0528314c9b9a118290de69dd
107525 F20110330_AABQYK gezan_s_Page_093.jp2
884f51446781c50028433008775b197e
fac282a56f4b1847d4b3d1b859f14d9c852a7d80
30266 F20110330_AABQXW gezan_s_Page_111.pro
223805264ed6c4d6f249b4f2e676ec9a
64ac3fa2383f4ede53462c277fb4be7de3265271
101620 F20110330_AABREF gezan_s_Page_041.jp2
d59ca09f8b02a796fc3ff02c2f004945
f462b59263d7d74a7632468379fc4269b50c5557
5389 F20110330_AABRDR gezan_s_Page_042.QC.jpg
af529833f2d0a39fe46f7bc390ba1be7
de1a48193e8e0c3c7a99a1beb9fdf8b5e208d571
23812 F20110330_AABQYL gezan_s_Page_012.pro
cfa253170031108627225612cb4fccbd
d2b09fa6e6a0f8eabea88ab53b0e40c7c4f859a1
1898 F20110330_AABQXX gezan_s_Page_069.txt
26aea70badb0b4d97f5daf63b8c71b7a
17adccdae9fd7cebc9363832109c85287b9abb3d
F20110330_AABREG gezan_s_Page_131.tif
e3642a0b722d19b45a5d1ef21e2f5a57
07bc6e26c6a4280d66316761b95c3b835aca53ab
52555 F20110330_AABQZA gezan_s_Page_017.pro
12590006184d0b8a2e48023dedab7497
b00f136250251cf28ac8ac552175006085c3744f
4790 F20110330_AABRDS gezan_s_Page_098thm.jpg
110c6e6c29881f286420e9ce6429cf2f
b0a8aeb45109765afe81129401218e38629411ae
76256 F20110330_AABQYM gezan_s_Page_075.jpg
1b1d6c0f667d6ed3665a102e5b194acf
8e876110e2d38260d30f57586edf3a42632eb08d
20819 F20110330_AABQXY gezan_s_Page_062.QC.jpg
d94a560793428c25ce53119bdb1c2446
646b5f9eb7667ddbd5c519fdd926dc54f6181c0e
74732 F20110330_AABREH gezan_s_Page_016.jpg
ec95801f3db6ef8d862a118113d81129
0fac2f186244d8ed73d8f381a47f9cadcffdc768
F20110330_AABQZB gezan_s_Page_019.tif
16e68ab72bccb1043d34b797a0483a1d
85a9a9754d929972ce9caa752c6a64163b9b04c5
77798 F20110330_AABRDT gezan_s_Page_132.jp2
13cf04b2a86739ec460811f8e63ab76d
bcd7f5c68ffeceed039ed5a57b22504db6790da3
5721 F20110330_AABQYN gezan_s_Page_099thm.jpg
6fb5c4a2d7e69ad998474fd6fd0485d4
8f4217f8854173a09a1646bd464440fbdbc33776
50801 F20110330_AABQXZ gezan_s_Page_084.pro
588dbd3dcd75532bde57dedb63e33de8
2bc25eba2dd02e77e6c855597a57a8bc92cc0662
5888 F20110330_AABREI gezan_s_Page_077thm.jpg
b7a3e8ee3dc99ff855da493469a1ebd1
f4c64f0897ba63f52163b05788d4d3612909b40f
2719 F20110330_AABQZC gezan_s_Page_010.txt
4991c28886a2b56e47a67364d928527c
ddcabb555866017096e153c67edc1f236b2f7a00
1572 F20110330_AABRDU gezan_s_Page_014.txt
2c7ffdf7c12b0d01801116734d90fa72
2553ae0e8d2ff693833f806e71ce098566cc59d5
2167 F20110330_AABREJ gezan_s_Page_037.txt
dfb97c92935c96425c64f2342ca42177
2cb20e2a6c63498ab136e7dfa63af703dc67fedd
93402 F20110330_AABQZD gezan_s_Page_078.jp2
b1bbb58448f79e3377a0add1aaad8426
c32b720357fec021d2a42e42a6c87ced415c1b01
4024 F20110330_AABRDV gezan_s_Page_052thm.jpg
972de2a871a694dab0fb5cc4ee1c6feb
baa99b3b475d3d6d09609ccc60ec8df1f5a72c05
F20110330_AABQYO gezan_s_Page_067.tif
308d8b576a0a2b6bbfd7401a13b6331c
7545e43b6c97c0ad9bf9a52139f7d7da554e8daa
76458 F20110330_AABREK gezan_s_Page_047.jpg
37e8bc8a7a48407642bb022915e1e5f7
c90022ba8355105f3f65d2d3be99c6aee073fc09
8731 F20110330_AABQZE gezan_s_Page_125.QC.jpg
472d069f6f3fad378b0f51b0d6e585a0
4613fd2d5c16a629563b5ae5fd3ab5331a326022
84988 F20110330_AABRDW gezan_s_Page_115.jp2
c3c5224a2c8fa5bd26d0e53e74142167
d420a570811aa476b9118959f623ab4c1ce15d00
83574 F20110330_AABQYP gezan_s_Page_084.jpg
dd20c973511943b8719df2416c98b611
bf76273ba22fbcb4c05dd1adc48513a5f13bf049
F20110330_AABREL gezan_s_Page_128.tif
acdbd086caffe30c53435bd53b939ffe
d95cfe62a5e9351867e47b42f297eb90c403bbda
92753 F20110330_AABQZF gezan_s_Page_043.jp2
394f9957cd57a5a3fa418e52975c3fd6
dcb918829003a4f3e72e648e51e46ebd41e956fa
51318 F20110330_AABRDX gezan_s_Page_070.pro
138f37560bc6be0b02138a45babf0bb9
fe0baa3f84c6452ce0c38a2b0e10b9dcf65735d8
5652 F20110330_AABQYQ gezan_s_Page_019thm.jpg
71a3573411d66064831f112b37caea88
234927ac4b85394232fca5555672fafde4655420
F20110330_AABRFA gezan_s_Page_039.tif
8384342279a6cab86049ff0c5a281c2b
a43a801bb519d8bb5a6164baf0ca0b880100bdc9
F20110330_AABREM gezan_s_Page_093.tif
eac4fcb7d2fded1b98ac48bf9ef9ede0
188911e2562286d97c881ad966274444e7f70170
4160 F20110330_AABQZG gezan_s_Page_123thm.jpg
9d31ed67dc693f50eb30afece3945780
30b1696f5b3700e025985bade43dc91380f99abd
97970 F20110330_AABQYR gezan_s_Page_070.jp2
04740a4bfbb6032302f402a188cbd0ce
77eb43bb834715c935c16d7651b986f63fbf19e8
49857 F20110330_AABRFB gezan_s_Page_057.pro
a1a3d68741425d128597e6c00730e11b
d9476b39e8b0a7d4d2749826048cdd0aa77c830c
21416 F20110330_AABREN gezan_s_Page_091.QC.jpg
7b348572cb3a8f3aebd6cd31f55b2d77
dbf303b97ffc554d2516b0e8de7f24119429cec6
F20110330_AABQZH gezan_s_Page_083.jp2
7ef7b393bf1b40d99241b279721f075b
f5396444bad9a2372aee868678d936a311ad56d9
F20110330_AABRDY gezan_s_Page_104.tif
c3cb52054cbaf0eaefecd4e2bc0871a5
155bdf116b7f7088e725c1d48dbf6ceafc568255
100574 F20110330_AABQYS gezan_s_Page_058.jp2
8764ed77c0d5e0565c48d23961ee5f5b
ba27130409138e2ce8bacb8777eb5dd81594c804
1930 F20110330_AABREO gezan_s_Page_085.txt
b08188ed1dd8f9e0aa87600ddbf3a36b
7de5014e3b50ef162e7b6f0cc017813ace1fd2cb
1992 F20110330_AABQZI gezan_s_Page_086.txt
429e3e5dc431bebc09578668e805d0e5
f62eecbd89f5676184a0e2284edb0320cf74aeba
17891 F20110330_AABRDZ gezan_s_Page_098.QC.jpg
ed73abf9a293643b8edd71d8a4740bbe
95ec280fa7219275ef6145eee7f97467e01e1fbd
51297 F20110330_AABQYT gezan_s_Page_083.pro
2c82970b9391d09a167e6ad67b6d9f06
a9e507a26ac3fbd214250fd5ccb5232acb5ea394
92612 F20110330_AABRFC gezan_s_Page_077.jp2
6a9ef012fa6b76d0f40b1f7121498845
f2f6ee7022ce45049f444f07b79f61c73ec91602
38328 F20110330_AABREP gezan_s_Page_118.pro
ab21672033fdd3e4f56a8717d73402d4
2a8b5fa5a7786c8c92478121cf67ad2236b5e218
6358 F20110330_AABQZJ gezan_s_Page_092thm.jpg
010aa06925a1c007c47c3f91e21e1ffe
105108bd94a28b57139ef1c89f6270657e4318d0
6274 F20110330_AABQYU gezan_s_Page_027thm.jpg
2ec42705baa588e59b624a4331db1d52
a439700ae58911f78bdc2a744706c23344aeaf04
5732 F20110330_AABRFD gezan_s_Page_071thm.jpg
b8ac9e158748e38be864b4ba66c6ef38
667cd2ffd217cdb6249289e6695403d7b6a88f1d
F20110330_AABREQ gezan_s_Page_060.tif
96f394d0a4faf6a2df6478d0cb49917e
db4ab6d70c6f02902c49f949d67b984cc0104eed
3389 F20110330_AABQZK gezan_s_Page_099.txt
0a8c264e747dd26b8106f9e9f2edc273
717e9bd1628b34f710c31d17a97c276352a39d3a
55001 F20110330_AABQYV gezan_s_Page_115.pro
971ed68f5b856670afbf96e73a01765a
6e69d7172d227f0e5e3247604056d26939162ff2
81301 F20110330_AABRFE gezan_s_Page_088.jpg
c884a4aacf866a14e9d977be77667637
3f561ccaad598d915f39e8699fc841e8da591438
18406 F20110330_AABRER gezan_s_Page_118.QC.jpg
bec64fbb79dc0796a46a3aadb8a744eb
c35491cc2e64ca490e764ef37364c89f1e12d552
18870 F20110330_AABQZL gezan_s_Page_060.QC.jpg
4312582f3b3f395959a352bf3f6760e9
65b1b491940c502ee60f6d372396146062ea3c25
4037 F20110330_AABQYW gezan_s_Page_072thm.jpg
ee2a5eeca9d2229a9432f69ed71a9e73
df05e0142e4c934922c5774955364e763de01616
19772 F20110330_AABRFF gezan_s_Page_114.QC.jpg
9b1a12107de2660382e16601426c8d00
47effd343a97e5f2df73cbd1a969e23f4ca58102
891684 F20110330_AABRES gezan_s_Page_014.jp2
febb1e41aac6bac87088eeceef024361
2e55c103a4d23b86cb3d2b20da3a1c69745452e9
4719 F20110330_AABQZM gezan_s_Page_081thm.jpg
d00b2b2f172fc6308d4f38a9cccf5ec7
449863665b863b46733a666eaff0124fa9af62b7
440 F20110330_AABQYX gezan_s_Page_001.txt
64505b18616e4bf4fed1eabc6414f90b
5a96809fcd97168a439dc43c537cd3f40b437e9c
86222 F20110330_AABRFG gezan_s_Page_089.jpg
6ec795a201ae3808fdb9004b8f535871
75514834fae6d3b9cf6dff1b05941c44bf8749c2
64902 F20110330_AABRET gezan_s_Page_078.pro
da8c24287c5584107190b7e37fed5e99
f05acdc522601ececcb5755c8fe4d6b15b3812be
39376 F20110330_AABQZN gezan_s_Page_100.pro
0f26ef515c90f2ae786b6b321b0d0713
09775e8eec575c113a288cfb6829c8fff2648190
F20110330_AABQYY gezan_s_Page_008.tif
bb836d049c7582c05c5a8aff64121e51
71be3d49cddb5769df700efb1b2665222ceec8cd
22742 F20110330_AABRFH gezan_s_Page_016.QC.jpg
a51ead6a45589e0cf73244ec3788e8bb
791698e537c93313f2ba7bd9a9dcd9c3b770e1ff
77835 F20110330_AABREU gezan_s_Page_027.jpg
c8d62142b28c4010b6a15aa4cb1d79a3
8a71eea249d3e64842f060ded933ab1202c0cd56
77431 F20110330_AABQZO gezan_s_Page_008.pro
51eeba6939e42a5bd94396a8a71193de
4fa02d473db6a00fb69439a11d720e1b9654b298
54965 F20110330_AABQYZ gezan_s_Page_032.jpg
975210cb15d71523b988434cd22516d7
bb7ac766bd9ff81a571256454fc2552f8888782a
106244 F20110330_AABRFI gezan_s_Page_079.jp2
9607e546e12412dfff667a8aa1af84d1
f03942b81bd815b6b42b36f0f6366709ac8d5157
F20110330_AABREV gezan_s_Page_056.tif
76a741aae8f85c50a0d2e98c2a3e3375
8335366e496e5b21735826d9b3573c33724c2c6c
81622 F20110330_AABRFJ gezan_s_Page_018.jpg
a72b5e679c2260b6f729613ad9cd78d5
217b4f6a9d06a451e1ce77814a7bfc0dd89d9e3e
2008 F20110330_AABREW gezan_s_Page_046.txt
f571667e9ff1657a63f96e0f620961e4
cbde87908763278a5c7a990c70415b1efae52bf9
F20110330_AABQZP gezan_s_Page_066.tif
2680fa66782bd02a720f09342c6a2d41
4a4861baf959d575da1c7960520284a27cee444f
71868 F20110330_AABRFK gezan_s_Page_021.jpg
f11411efd3e583949c7ee432629b96dc
18812a9c2bb6f28fa52dccaae93b8abaa06187aa
F20110330_AABREX gezan_s_Page_065.tif
39af3809269d876079cb26ef56538e71
465f261f4012c43176d76aa5581e24c378b04f95
87087 F20110330_AABQZQ gezan_s_Page_006.jpg
b8ae1f5f31ff28d3c73124ea52e32d88
14d3fcd0b47ce8adf1aced17df1900e50be1e953
F20110330_AABRFL gezan_s_Page_119.tif
f19a6bdc684dda10aaf32da0798cc1e1
3f9672a877811da341b853f833152eff4a93700d
69229 F20110330_AABREY gezan_s_Page_091.jpg
bed94b179ccdfe681ea1579c351ce16d
3aed4b89c3f51fdc74db751d9f9c64daeda43eca
85862 F20110330_AABQZR gezan_s_Page_107.jp2
45fe31d77fa52b39656f6ae7d6eac084
7707bf1d9484797fe72294d96ec83a6f4e61e54e
2342 F20110330_AABRGA gezan_s_Page_128.txt
59267380daf1a2c9cc531697ebd84aa3
5cd3488e2c75fc624c6afb30e07022882e839285
13278 F20110330_AABRFM gezan_s_Page_124.QC.jpg
e7d8c67c5f00402e654f28b27af48cfe
ed11aed3ef286c38e5d689fe1ed5827b28219455
111493 F20110330_AABQZS gezan_s_Page_010.jpg
c8209f5b6ff2f0107a46bb6f541b0a89
978449dd5f2e2f1fb310168bf3f31f23e9f90c41
80893 F20110330_AABRGB gezan_s_Page_090.jpg
aebe7f28cafdacfaf2fba04a5d5f7f39
a8e0f2ab903015f24898bfc49ad15e5f33b9def5
105445 F20110330_AABRFN gezan_s_Page_117.jp2
66c8e177b3b35b83f325cce13e6f9e36
cb05a8f154e8956bfb8dce7c2807746d04a05fea
1844 F20110330_AABREZ gezan_s_Page_019.txt
758addf4ef263d887bcd02fb413f0e28
4b97f5cb748c509cd428609182f63ee63af85855
1183 F20110330_AABQZT gezan_s_Page_125.txt
a0c43189fbf5d1ec1b228d27bc21b752
a553ebf3bd0336c695b08a3005970f5832899b52
24909 F20110330_AABRGC gezan_s_Page_069.QC.jpg
17c249a8e5670b557df4456daf07edae
c854b6e41dfe170b60b4a4e870105d8f8c5a8623
30660 F20110330_AABRFO gezan_s_Page_011.QC.jpg
b846d223a271e4b3210377bca99644c0
c88088ffe0c7878bed02edfcf958b62436fe40b3
16414 F20110330_AABQZU gezan_s_Page_072.QC.jpg
0b719892409454b2fc404e55520c3f7e
cd00ee7d77852eafb64bc89e12572297fa5fcbd0
48823 F20110330_AABRGD gezan_s_Page_090.pro
a277be578db7b61e5e1e0061da61e597
fbfb630ae276143ebc4c44b1f59fb70a0404573d
1977 F20110330_AABRFP gezan_s_Page_027.txt
a4f15869713c1a35ee2aa0f84f9e6c07
5e0e04268d1ed816a1ea1c1abf1ebfd1f1fe0e67
22028 F20110330_AABQZV gezan_s_Page_021.QC.jpg
393f556a65d09f25882587c0b2c9b117
e56e02a30b0f241fd94c0020926a0b7115dfbcff
68329 F20110330_AABRGE gezan_s_Page_059.jp2
bfc4ebfc878e5b5e1dddbefdd8ff0171
68f7e11fc7889834ecba5c53058481d452f11f0c
83300 F20110330_AABRFQ gezan_s_Page_050.jpg
23b2424faf293d27f1a4a9becb120db2
4077ea74462da8e9f8dff07d9ae10a92524d31b6
82050 F20110330_AABQZW gezan_s_Page_119.jpg
04e133c44b5a921ed86c03617dc4e0c0
f35228b44830c8a4a4a4001bc03d02bc9103e5bd
5592 F20110330_AABRGF gezan_s_Page_110thm.jpg
be934748c43078bf48a37a86fd81b5dc
427a66e17c249199c66d0b89ae7c64acdcfbdb79
F20110330_AABRFR gezan_s_Page_070.tif
3ebfe3c8cc45b9381b3afd619673e7bc
3af25c16822bcbe6d9fb56eaea096bb16305a039
F20110330_AABQZX gezan_s_Page_115.tif
b843bff101e8f05e414292be8afac81b
6c87049d9b1fc37b2911ce732eeb4c58b0a8967a
27034 F20110330_AABRGG gezan_s_Page_055.QC.jpg
66973709a1b63a3fd04286919474d561
41d7d26b52a31dba8329ffec3cb66f2d79586414
3183 F20110330_AABRFS gezan_s_Page_011.txt
4f6d286e617f664ba5879cc356e42729
b70db978b24e1b7aedfe7f76d5e26a68fa1a0128
5995 F20110330_AABQZY gezan_s_Page_074thm.jpg
1d0cc4ce834e8aa12d63fbd85356e2e9
e83e77e8e7d8fcd39fcb6e56f6cf9bf45d5f1784
1184 F20110330_AABRGH gezan_s_Page_060.txt
c9b7f614ae0ec104ddd90d54d9600790
c921109b9ea911808f67ea20db4d96d0df03f9c7
633581 F20110330_AABRFT gezan_s_Page_081.jp2
e8ff169fdc7b93af81555cb682db0c70
922d984c5435199884235924469578b1622815eb
6203 F20110330_AABQZZ gezan_s_Page_086thm.jpg
f46dfca812bbacf142da7ca9512b24a3
c2e991f93624c92a7bab218c51b189447e48cee1
53215 F20110330_AABRGI gezan_s_Page_089.pro
53392b0a00ec368aeaa98e4ad27097cc
ea11ef80832d6244e20e1369207fdf929df52bf0
F20110330_AABRFU gezan_s_Page_018.tif
e1d58fd0e6a7b579c1ae80395c73fa63
56ab32a9f54de1d1f41b81d29ad42106e2c5122b
104212 F20110330_AABRGJ gezan_s_Page_045.jp2
9b9b4f3ed5cadc9e1f20869073ffc1a9
ccb4afd1769610e2f7e69ba302128d09a175651d
F20110330_AABRFV gezan_s_Page_006.tif
2fdf01856e4a3e4e7054262c1dfbdb9e
4d02926194c669241f60f2e2e63a2dddcd91b092
1407 F20110330_AABRGK gezan_s_Page_002.pro
b95a39f89ea2bf2ceeceab6268fae439
802707545d2604f3fc4bd14444337f326946994f
2930 F20110330_AABRFW gezan_s_Page_076.txt
dca7335c27021ffb2e84a7741f9b36d2
1c540e9308835fa1596c8529d2e99025ea39556c
20227 F20110330_AABRGL gezan_s_Page_095.QC.jpg
dde9088e5fd9e98dae012908a5b56713
24f44e25c55c0a820fe90969bd9279d535988349
1967 F20110330_AABRFX gezan_s_Page_044.txt
4cde86e27785ff68804a512deb90d5e1
14f149687aaabe00084e28766b4060885e6336c2
83576 F20110330_AABRHA gezan_s_Page_092.jpg
4d9ca044db536540049f9df340d98ed2
e8a5181dada8297b9096ba0121c42b29c34b3342
6402 F20110330_AABRGM gezan_s_Page_017thm.jpg
d0b87a6b006ba4ccdb31f1482901e80f
d0ddc052b05b88e1f715e498d6318cc5a4da3ec6
97376 F20110330_AABRFY gezan_s_Page_067.jp2
1a566bd3d126c229eb5afb608ef3a369
4163c2c86b6522d96154b888eb1b351347f97012
1911 F20110330_AABRHB gezan_s_Page_109.txt
34ebe91a0a11b5bd2ebba5ceec823b0e
f5259949b69f27acbdea5483d5d6291df8072a0d
69745 F20110330_AABRGN gezan_s_Page_123.jpg
3962587ee6a229f6121ad2e8fa577ba6
b117ce0459be428301161ea2a0add1cbbcaa6644
5117 F20110330_AABRFZ gezan_s_Page_112thm.jpg
737c242ca77491ac01b0fce1cb898737
2abaa48f5c811db16b1fbb62e3d17311d72329e5
25138 F20110330_AABRHC gezan_s_Page_080.QC.jpg
bd233c3ba20cd042cbae35d3d860af1e
8856540e4b17792222f9f5319e8f5084f0aea97b
64747 F20110330_AABRGO gezan_s_Page_062.jpg
4416400cd39a5e5d4417b7e43c301c81
26c7aca1069bdfd8def4febfd8b972b045947edf
26734 F20110330_AABRHD gezan_s_Page_036.QC.jpg
514f4e65bccd26cc290dbcfa9c9981a3
dcd03713d6204468a7d8f7c754354e2c88ea3492
6005 F20110330_AABRGP gezan_s_Page_065thm.jpg
b143ff82976c2919236188e1d4f52f53
1f663cd4cf94a93a7288a2abd93a21ed60eac56b
22899 F20110330_AABRHE gezan_s_Page_119.QC.jpg
10f28e00581342c8fed6dacf6142fc4f
0befb9f65434494661d86dd0af20d23d150bb14f
46475 F20110330_AABRGQ gezan_s_Page_019.pro
549ac134e66311ea55e369bce599348d
996688759d3936ffd1299db2e5d6163ef0874524
5830 F20110330_AABRHF gezan_s_Page_078thm.jpg
779df12f5643250ea8b0f18b55bb8452
64848ca71c050637512225d3b75cc2c344b807a5
F20110330_AABRGR gezan_s_Page_115thm.jpg
3c859c5414eed7312733df6433917e3d
77e98c75a5d9490a5b199258c0ee7f9219f888ba
1140 F20110330_AABRHG gezan_s_Page_121.txt
d5f80d3d86d2a3613477c77272ee84bc
3efee5d7ed4a69150e9d0e8549cad315d23aed99
F20110330_AABRGS gezan_s_Page_090thm.jpg
3c9af0f5ff5de1f2ab7f33bedb7bfa2b
80b2b388aa61a823a325603589ec0f52752398d5
2393 F20110330_AABRHH gezan_s_Page_129.txt
424f3a90607a243f990985037541da19
bb14ecdc48b9b9047b20ef0745728250f2361e59
1051978 F20110330_AABRGT gezan_s_Page_090.jp2
30352fc3824691bbb492a9c468054303
9efc218675cf247b1b7273811c8ed869cecf712f
1351 F20110330_AABRHI gezan_s_Page_120.txt
a99d7dca2169878a1bfdcce4293ddc25
4f4189547ca8d1179b26454ed141aba774b71aa0
14241 F20110330_AABRGU gezan_s_Page_052.QC.jpg
3dea7e6beeffbbdc3c16f7c21e198c48
a14d7327c85d7a68170add97de0f96f026f5b87d
80606 F20110330_AABRHJ gezan_s_Page_112.jp2
f685eb3041ece4c84fa868e818053d31
d8b51e7de8ec4ee403f84c96f09efad87ab74344
23766 F20110330_AABRGV gezan_s_Page_019.QC.jpg
0ad299fe9da690933e9d1a779987b62b
4c93c9d569e9c24e9e2236266d57931fe5e1bcab
50527 F20110330_AABRHK gezan_s_Page_024.pro
74feb37b9959f2b80e06e7e9dd26d0fb
573b20f6f2686b67e0f8101695177da090f83a34
16437 F20110330_AABRGW gezan_s_Page_032.QC.jpg
5585990f785d50975745ad231bd83a73
2dc4b94be07dce77f63b2c860774fdf6100dba1e
84748 F20110330_AABRHL gezan_s_Page_057.jpg
90f8529acf2966955e16abdc8f5aceb1
37725303ea2ce1a4e835c8103c4dae14ee3425eb
1927 F20110330_AABRGX gezan_s_Page_029.txt
de52fa9fe13a6abf0d111011b4b74783
7ddae79d348e5ede4ac34b1671a2e6a81627e1a9
4344 F20110330_AABRHM gezan_s_Page_103thm.jpg
3cddeffed4ac389e6bf26a6c91c1e9ab
501ec62747c2e098102e040c316ac8239056c6c6
F20110330_AABRGY gezan_s_Page_130.tif
3cded78c97912bb22bd136c3aef9aa56
6345e218532f5f89a2eaa9fdcb0715a5e15c9fc2
1793 F20110330_AABRIA gezan_s_Page_016.txt
99f18d0bc5ec862d99aacaa1f95d8082
e245d6cd916eea76ef67cd617e3466cb06121cc5
880257 F20110330_AABRHN gezan_s_Page_013.jp2
c9283af2992048b1f557f6d841b01643
68b83a88712026e2775ee403da8af05cc5daebd5
1051962 F20110330_AABRGZ gezan_s_Page_008.jp2
82de38ae7388dfda74808ff9a8b1ba1e
adcb3bf1579e935ddc422cdda5b17350c2634987
F20110330_AABRIB gezan_s_Page_005.tif
a28fe4103942b50887d010ba350d00d7
45c6eb7fa1184d42d211164a17d90113bbfe274c
77148 F20110330_AABRHO gezan_s_Page_061.jpg
04316ed9898761dc6fd24315dce841f7
02c029057e08c5058703cadbca97f6cabe860326
2475 F20110330_AABRIC gezan_s_Page_059.txt
46d6623f32ab173555c7ba9ae97ee0c7
2ec2c87656afc29ac67afeb514856371d2835257
21521 F20110330_AABRHP gezan_s_Page_042.jp2
99962ff2c3b0e8a4021f01e3bd4d9361
39eca59e59d622d03fdd3f3b660a16ab7bb65a1c
2004 F20110330_AABRID gezan_s_Page_040.txt
b174cfb2d2dee2c135df979993aefba6
08d270548c7fb95c332e40a41f47f232df9acb1a
F20110330_AABRHQ gezan_s_Page_011.jp2
7c6ac1d2c407e2ad27ad4deb783e5e88
91da3d881000752c5c2bc7c175766c3a807d9d5a
F20110330_AABRIE gezan_s_Page_012.tif
446dc0577388f3b546b5c3b6ea311212
a591f47012eefc1613f8c77385087970be2bea49
77926 F20110330_AABRHR gezan_s_Page_070.jpg
251371c3faa2d3d7335eb3d87b825345
8e3572e20a5b15cc039acf2b5bde2abcc9728f51
2440 F20110330_AABRIF gezan_s_Page_047.txt
6cee5f668c0b8e6ac24b81bdbcda3cb6
ad1a1b23b2ca3a606a0052db2a349eeb9ffcd508
3076 F20110330_AABRHS gezan_s_Page_116.txt
90b25bf2b5ee6ebc47aa01555a8225aa
a54c0653f347c09f553093a1457e9d307c9f6357
46443 F20110330_AABRIG gezan_s_Page_015.pro
cc3dad1d6c5b70d79415b2ba2227416f
24ed794562a9021103b5a44c350ae4f0397b5b39
2012 F20110330_AABRHT gezan_s_Page_036.txt
0d9a1391fd6f1f22bcafd16d71ee5a40
ddec61dbc44cb7607a615d361ccaa60353c8e911
848309 F20110330_AABRIH gezan_s_Page_100.jp2
219faefbfd5637818b2fd0445f8da74b
0a68b09be81fc9fea79e5b509c33f57a68c99cba
F20110330_AABRHU gezan_s_Page_069.tif
d3bfaa6a9141f9723740d6df468ee9d1
0f46b703ba01132f3d87720f99818726675b7104
F20110330_AABRII gezan_s_Page_042.tif
bcd2a53d2130e925abe9d06f41fd69a8
ee5c0f558ace919e7e4ddced8659599d3b9adbfc
101986 F20110330_AABRHV gezan_s_Page_054.jp2
8612e6efb296897961413514c92433b7
bbbd2bb11e03758cfb57578693bdc524ed1724a2
25105 F20110330_AABRIJ gezan_s_Page_033.QC.jpg
2973aeded8efdc519433f2d047021e02
0a76148545168035f193202a463572e74b3fa180
6300 F20110330_AABRHW gezan_s_Page_035thm.jpg
a0fe7f5aa09032b495231685511aee09
359d34ca87e5cb2128290f016feaff31d5e10faa
1873 F20110330_AABRIK gezan_s_Page_054.txt
04fe9b79cef9652c4daac66429bc501f
60104cb67cb353a1b3de0d94ee715b0080ae37d7
6275 F20110330_AABRHX gezan_s_Page_084thm.jpg
3218a4d054069eb8416a8568933a70e4
b75c0b6d8878df57dca71bb64ea746f337d64cab
67598 F20110330_AABRIL gezan_s_Page_100.jpg
81055d8712846137feee29e708a5451f
b4e71dba98cca9b562ee23c4c005c6a5371e72a4
106203 F20110330_AABRHY gezan_s_Page_105.jp2
93920b45164df2699ae0302a31f7f98c
0ca68de78f583dae296d1ce57c45a9ee82ba931e
79174 F20110330_AABRJA gezan_s_Page_108.jpg
43186e4af7896ad9da86031cf1d01198
eb17342266ed0ee28b3427f668e725c8c99b07b3
24731 F20110330_AABRIM gezan_s_Page_058.QC.jpg
c6aec44ebea22f07967562c3bbe1f826
d0eee357e9009ef1ddba4a8238c6fa13e1794b7a
30832 F20110330_AABRHZ gezan_s_Page_008.QC.jpg
19657fc5c3eba7da6ea5801358a33109
6518e6556dec113147bf5a0618727679c3d5d5f3
6019 F20110330_AABRJB gezan_s_Page_073thm.jpg
fb6eef1f6ce63afb3c852f516133f165
1f524cd596b967e22752be74a87c4d64943fb064
4181 F20110330_AABRIN gezan_s_Page_106thm.jpg
33f6fba8af2751ca4f23b305ece9dfa2
85f875d65a97b381388dea67bb3a9106d1082866
65383 F20110330_AABRJC gezan_s_Page_107.jpg
c9396ef22b91dc2b91d723ad8fecdf42
e1ff46d1ce55062aadb425d86fe16f7868337de7
50214 F20110330_AABRIO gezan_s_Page_050.pro
b8a8df5b756d1f9ce6396992fe9c34df
0f9739dedf0a0a24f1e8ea52960d90c8c4fbe716
2142 F20110330_AABRJD gezan_s_Page_055.txt
b2934482e5fd700e81e65301a3606c4c
3a54f87cf41b05c55a743a54ae3a1a25dc37c949
F20110330_AABRIP gezan_s_Page_109thm.jpg
b5844ce2ad404008fce502384b7a4364
9deee233956b17c0c2c23e51e84d6d61600c26de
5882 F20110330_AABRJE gezan_s_Page_034thm.jpg
6f17f5dc3345db10ee9a4088c6112009
3d20c4105a4a7aafcb9f6a21fb98b8ec36d0e07b
26052 F20110330_AABRIQ gezan_s_Page_084.QC.jpg
dc4e06b618e4a5bb3230f4f2c9aae32a
a14d42ccd8390a8916999efb2f32a25223330992
F20110330_AABRJF gezan_s_Page_034.tif
3ff7ece63dfe9d4bfa1b6d17959503e4
8b91f7aafb11b0838611193c337d60bdc7e166cc
3055 F20110330_AABRIR gezan_s_Page_009.txt
1262190e1b30553cde4a646a992f8d1e
c415e0ac9d90d55e8e05fd0de43d1437f0ff0a95
5678 F20110330_AABRJG gezan_s_Page_058thm.jpg
c1c709b659c6376c451d94a72d82e8ed
6a1d52b960c686910086cf71ce40ac300739247b
39143 F20110330_AABRIS gezan_s_Page_107.pro
dae61291a80adb31446564b56016d1c8
a3eeefde86b59d411a810d70379ffca7b53cd823
21789 F20110330_AABRJH gezan_s_Page_113.QC.jpg
a3cb3023e85274d8c1510e3e41883856
30e9b03051e413628274c9400d259b40a33f9617
86678 F20110330_AABRIT gezan_s_Page_087.jpg
797acbe83117246c698aabf73248db24
e840898639f5c2f09218ad5ad4e33b4f7ecf0b34
57809 F20110330_AABRJI gezan_s_Page_128.pro
613a49eb0bce59e99eafcaae58a11f20
fa6053e9df0fc8fa8337635df28653312dd52d4f
F20110330_AABRIU gezan_s_Page_043.tif
2f654938c44cbbd968c40cedfe514285
27f4bea1a38b48618f7c54fe8377453d7a1b9228
25407 F20110330_AABRJJ gezan_s_Page_030.QC.jpg
aeb3430f268a07bd510fb59da7f8b727
daa99f27661cfe1e86e964ea4532efd001c97e50
6539 F20110330_AABRIV gezan_s_Page_128thm.jpg
30fd33d29e25df338a8b91a199372cf2
056b0fda7a8d365652e5e325b8170cf790e279df
102763 F20110330_AABRJK gezan_s_Page_094.jp2
73815c54933cbee33b7db3ffcaaae36f
07c0efb520657fed2d6e6debc139f1490591cf02
F20110330_AABRIW gezan_s_Page_059.tif
9a980d58815d747fe98434f904894940
5257bd1d1301ff573b7c61638ce3a4691928d758
5996 F20110330_AABRJL gezan_s_Page_116thm.jpg
80743395e0efddc838821dd66f9b49be
e4e6f3a38d346902eec9b8c61ff9b66af72e3edd
31896 F20110330_AABRIX gezan_s_Page_103.pro
8106c91320409f771da9eae86555c7d8
76bc20acc8abb49b98dd2b1ede88c7b1aa33b606
24677 F20110330_AABRJM gezan_s_Page_088.QC.jpg
7f2b6048bb5d443a960d3c00d39bf170
f860c4f16b5f2522526574b5faa0641109afea8b
24818 F20110330_AABRIY gezan_s_Page_039.pro
9495851c0036314e3bbc007fb0f4b820
e098b862249c2a42bf0eb342ba812ef4c6c14f80
98593 F20110330_AABRKA gezan_s_Page_016.jp2
468c3398a8bd461c4cf45a0d8ea0df7c
1cb23a0cbea690a30a4c8809608728f4f8ca923c
51993 F20110330_AABRJN gezan_s_Page_104.pro
a694691d364bfa60f7dca5d3c5470820
732c34a35166cdfb01eabb0313cfa3965d71eff7
2068 F20110330_AABRIZ gezan_s_Page_064.txt
b9485ef891621cb681c5756e2a6dc523
0ba4f640b1c9a96847d66875e69a4a1336349783
22780 F20110330_AABRKB gezan_s_Page_075.QC.jpg
12c9e0d3ef659a17922a3806624e4e83
4c349571a77a3c514ee91c7f059f192c5d142251
43126 F20110330_AABRJO gezan_s_Page_052.jpg
9b06de3a2d96e530a325763c0247866d
3428c193799abea753d18934a485d2cd20c5cb06
25761 F20110330_AABRKC gezan_s_Page_066.QC.jpg
f43f5dd6140426ac9f86403079d393a0
2e0e23c25f2d82a320ab233cc527858af364db69
57957 F20110330_AABRJP gezan_s_Page_117.pro
129af5ea08db36cfca57c85850d3be6d
e883e1bccc7e36a80c22e3a61c706d137feeb9d5
1788 F20110330_AABRKD gezan_s_Page_110.txt
65329c27a3918690f10249b033f06bca
8d065ec27ef8eabb5e0253a883cebf702729353b
496 F20110330_AABRJQ gezan_s_Page_131.txt
10e2396668fc82ede8abfd9325a3435e
8584ed35ec15695d5d8fc329c7efb96d773a1db4
102011 F20110330_AABRKE gezan_s_Page_007.jpg
5e12cb6a32483c34d1f0dea9f8ecbde9
589ac67c52206d84b371b4c577ccf0fdd900e78b
74035 F20110330_AABRJR gezan_s_Page_076.jpg
8f2082f2111b55de90afa7d7ef32a1ba
f57b86dd94e30b420ed42b6d17eb09f0c2dc91ec
F20110330_AABRKF gezan_s_Page_038.tif
e83750cdbdf5c7d70d604b1f69801610
75291b92cb15e679050a1745e4df069d32e2ce07
6163 F20110330_AABRJS gezan_s_Page_049thm.jpg
ffb2c5e5d7047ecbf7420223a015c1a1
d019cb68db37a494d42755ff43282a21a83c143f
4180 F20110330_AABRKG gezan_s_Page_005thm.jpg
32039c01ed912f9c2c5be7fe435c1bda
da95e9eb05edd65ec6a71f685eddff064249e763
712865 F20110330_AABRJT gezan_s_Page_103.jp2
2fb5fdd6496f24fac12677c4d6b5b001
725fe3cc1ab48e9a0dea52cdc6f018c775188232
75864 F20110330_AABRKH gezan_s_Page_015.jpg
549125eb571e7e06f46da3d4c050a7a3
62085db478447c7ecf07b6441a952c03e25fb783
1926 F20110330_AABRJU gezan_s_Page_018.txt
b40c56a6eb97b6a81ac191d8c0cbd7ae
ead37039ac405a63ab39cfe83438f5f79c89b81c
5847 F20110330_AABRKI gezan_s_Page_100thm.jpg
620376874dfe1d258281dd95886b8690
d99c6eaa6afd60876d9385427ef86b1197752bae
19822 F20110330_AABRJV gezan_s_Page_112.QC.jpg
8025ceca408fc5bbb9c11fd3cd14d465
73d6b454e03b8fd1dca742eef635bfa0a154b66b
98118 F20110330_AABRKJ gezan_s_Page_071.jp2
90278918b8e48eb771c86099b3ee3131
b827d3ddaf0e70be1f4c25ef65ab3950dcd05fcb
56926 F20110330_AABRJW gezan_s_Page_025.jpg
6f293bd973b9e1adb3c61361707741ce
e81830de36b2019a82c6ffe60e8a62ba1966b9ef
17318 F20110330_AABRKK gezan_s_Page_120.QC.jpg
8117f1af2af385cf6f0bfaf3d19450f8
c4583dbe7691e1edfca980e577e54093d6e3a078
52744 F20110330_AABRJX gezan_s_Page_064.pro
c85c6e950310b5ad4d4b7872fbf37406
ec4ef6529c269c9439b615c5b7e97672e0948310
44587 F20110330_AABRKL gezan_s_Page_031.pro
70feea88a489c62b422d1cfce481fe37
0073d8cf98a31d94f00017b9a5a1d32af9c71ece
42235 F20110330_AABRJY gezan_s_Page_043.pro
b89fbc59a233e7576058cc5a0f277568
f00eea5b3bb1a169dde9afb8191526c13f6fa63d
19301 F20110330_AABRLA gezan_s_Page_072.pro
caf47071f9461cefa319121d247e284f
13610dd123e0a9866a8ec47f245327fa2a0a0866
3044 F20110330_AABRKM gezan_s_Page_117.txt
c6215fb79d10b8016ef72ae57900a65a
80fcf7d13b801e680af691e0ab627491753ed5ab
24656 F20110330_AABRJZ gezan_s_Page_035.QC.jpg
16a40fc6a21c3afc2b4609f1cfe50033
2490adba73f5939ef9da46a0e1566a13d27fa24f
F20110330_AABRLB gezan_s_Page_049.tif
2ee0b7c22a4507fa58437eb77d247690
3b5df2c47c06ee5e463a3337097c052b15c61839
10600 F20110330_AABRKN gezan_s_Page_012.QC.jpg
da41c5b0b6b08be5529ece0c9afa511a
a3c3c692cbede7eddb4024ef0cb9cdc339cc221b
5285 F20110330_AABRLC gezan_s_Page_039thm.jpg
5157e3e1b5460cf7e5cab8296c7a4c1e
46f129a5176c3a0ab454325307f4507fbdfb8718
77629 F20110330_AABRKO gezan_s_Page_041.jpg
55c8cb597ea9ee9424b69308cea15adc
961577ba0fcc5dbb75f8c811d0cb76d1cad7ec58
102526 F20110330_AABRLD gezan_s_Page_074.jp2
d6a39e21c50b3529b72c341334f62f8d
a7bf7688010cba1450ee5ac6baa12a2cd644c37d
F20110330_AABRKP gezan_s_Page_051.tif
762d075d125eb69458d80b6b2cd426e2
6338554b900c27e34091ebfd9eb818834d15fac9
79751 F20110330_AABRLE gezan_s_Page_097.jpg
2cbdad6a26babab54b9945697a83b0b9
2c7dca1e98477d3b1f05fde18b7fec7d262e80e4
76448 F20110330_AABRKQ gezan_s_Page_077.jpg
3dd6367eff95d28997dcfe0a6c48787b
f1c8511186c9bd7b3ef6b618ea0a42f89ad86bd9
F20110330_AABRLF gezan_s_Page_020.tif
609be266bd6bcabc63b54a60539be3a5
31001a62fedc93abec1d13299f3cb89308594003
83791 F20110330_AABRKR gezan_s_Page_023.jpg
39f96ddeff5927d78d1a542972e29f32
a43e8098cd27362d0097a81eca75a610a5000c4c
1198 F20110330_AABRLG gezan_s_Page_039.txt
3b7368ac58db7adcee86e6274e110b9a
8d23133d57a585e82e4d7cb330d8bfa6ef58f21b
F20110330_AABRKS gezan_s_Page_014.tif
4c3cf89f16609dd05af33aafa72cd71e
3ffe733d12f301b161f30bb49a24549b8a795739
6299 F20110330_AABRLH gezan_s_Page_050thm.jpg
33f55576cd23b031144b89a1852e24cf
cf3b5c4ba599dacc43c2118b6e9594a89e07944c
6531 F20110330_AABRKT gezan_s_Page_127thm.jpg
a1db351ca38fb21911ab49d0eb4ae34e
8583c21528a85ccfd3fad0fdc5a0ba35b70f0a6c
79155 F20110330_AABRLI gezan_s_Page_048.jpg
1e07aeac1f3e3ffc95bab6f80054b791
9936e416d5c54a5d3d7df5809d9803371e1eee47
47437 F20110330_AABRKU gezan_s_Page_034.pro
b0c2c1385912ef21402dc917c4ea5ff4
3b87cf738ce8a3ccd870110eb99447fc1488833c
61983 F20110330_AABRLJ gezan_s_Page_004.jpg
784b6c24fad0f0a580e4059a08a967ba
58e03e505c4764551ef4f0167840b0e18e91dc5a
99846 F20110330_AABRKV gezan_s_Page_110.jp2
f3bce47dc892a71c1f7f5a9f84ff7b37
3b29f3e6d64e9efa32cc0a01ad8ba3b819ebf6ae
6268 F20110330_AABRLK gezan_s_Page_085thm.jpg
bf3c283bd4160f81c57ee37279abe370
31ebbdd182fa80516cd6044c27d8a360cbc922b5
60450 F20110330_AABRKW gezan_s_Page_132.jpg
ce3426410474f997e6c5772c64de080a
8cffb3c6f5f8e6799bc53f88c6b269e9369dfa0e
4979 F20110330_AABRMA gezan_s_Page_004thm.jpg
000a421ddeb37c92526075777796acab
f5ce08f6335263be1ab042ec0dc3f4de49f676e8
26922 F20110330_AABRLL gezan_s_Page_089.QC.jpg
3d909c12965341b5b98722f377fdbaf1
51708c4486f162899635d56949af3f72dbd641d1
F20110330_AABRKX gezan_s_Page_001.tif
08dcb56db10ecb298be5e17ee0d03f8b
fd8798cf76d3fcb281536964166b62093498f651
23806 F20110330_AABRLM gezan_s_Page_037.QC.jpg
88cee32cdb193272ba8695e14ab84a4b
1bc93613e80350a2396a360136702255fc506b13
85893 F20110330_AABRKY gezan_s_Page_036.jpg
df1049a91032567e70ae8f6c33836c9a
942e8d7b4b798b9e3159c0d533a588b6c66f8341
F20110330_AABRMB gezan_s_Page_094.tif
435b737be5be20259c7484b200d6b59f
c4fb8c317fab1fe9c1623633cfa73a5deec744e1
6196 F20110330_AABRLN gezan_s_Page_023thm.jpg
d8c5de95d7955ff8111e5e86bea0695f
9bb5291d35ca644f99668c8115bf6a15ebbe2e8f
56023 F20110330_AABRKZ gezan_s_Page_103.jpg
bb9cef2f8581b2c6d934d6aea594d1a0
861a94b39102949e18649461caf85131b2f24dc9
45696 F20110330_AABRMC gezan_s_Page_058.pro
d59f47d652a9886eccc4e9992c100190
8ede31e00091237020cc2881c442b8504e58ef92
17340 F20110330_AABRLO gezan_s_Page_005.QC.jpg
5faa95184a9f25e2f14e9fe912a64c4b
d4c6f3dd63e40a0016f42401a12ec9903620b23b
1714 F20110330_AABRMD gezan_s_Page_013.txt
cde42a5e5f717c12b27c510a2da2cbb0
7a40ee074c926b6366468861aa826c8e7c10f550
2052 F20110330_AABRLP gezan_s_Page_104.txt
9071439e88a3e4da3f1ff6741a559a74
51108627381b68b1a476e84b7272ba2e268635f2
F20110330_AABRME gezan_s_Page_055.tif
b945f06edca2695f756bd57b0cee0ab8
46de9005d808618087c8c4394d62cd3d492a334f
21071 F20110330_AABRLQ gezan_s_Page_125.pro
01bf46b8c912ada326f5653c138b2c10
c508ad50d0b816382d915cca7f0c660760d8f1c9
4742 F20110330_AABRMF gezan_s_Page_025thm.jpg
e25f521d3599d34d8b282f4487d8af5f
ed900a9dbca52427de05d965358afeb56cb4b756
24977 F20110330_AABRLR gezan_s_Page_046.QC.jpg
327b77ce9c70e60ab6bd328eb13bc78e
b88c82543fc6551c4743931ad77c0c3c3f2c7303
6873 F20110330_AABRMG gezan_s_Page_011thm.jpg
c1deda6b6681654b8381eb3e96872b94
cafc54b3530dc0fd72ac583b0c31faa511ec7d24
F20110330_AABRLS gezan_s_Page_037.tif
52d2680537be931d689e956599f36127
5e201296848358de34e5003b09bfbc5f8292f422
43802 F20110330_AABRMH gezan_s_Page_126.pro
1bd9c53621f0422455d6547129da6e75
6046f431d05ecad268a29a25f6fd85e2fc2f233c
F20110330_AABRLT gezan_s_Page_124.tif
eaed02d7d7af548df1e0f769b2d4040b
ca31d5ff49d3bf4174022bdfd29b5f2f443533bd
F20110330_AABRMI gezan_s_Page_023.txt
c8063d4509804c6cd17ad62c6d0d958a
9194cc21cc4cd2217fc9c64385aea0aed5f7ace9
F20110330_AABRLU gezan_s_Page_010.tif
29b5057779976496afc63cd3e3692a36
e8711e485346ecc62debab6730accefe39ccb4a7
73639 F20110330_AABRMJ gezan_s_Page_037.jpg
0c25b8f9549829b04766ad99bfbccd3b
20b1871b3f21d2dafe11dfa6cba53ad886dde835
52451 F20110330_AABRLV gezan_s_Page_055.pro
aa79ed3cb42f0cd4b9c0a80defeb57a2
8dede99eb07da7ca1a274132f54fe87063d6e678
F20110330_AABRMK gezan_s_Page_061.txt
ff64ee889695b379dd4eaf358c38de7f
227b84ee6f63452a77bc5e5870b2316cddeadbb0
76607 F20110330_AABRLW gezan_s_Page_019.jpg
5f2bd9385a1bf33069e27ce25f4fb984
52ff2b2a14c2134778ae1bace601d6be1ac886ed
4538 F20110330_AABRML gezan_s_Page_118thm.jpg
dd46bf232c2763dea535e50cec0ac1cf
99ecdcd8163d663bc30586ffa2b85e472fca58b2
5781 F20110330_AABRLX gezan_s_Page_007thm.jpg
bda411c7e87cfc69371d106c2fb69643
e8c744d07ad287bb48d6ac1be7eae46417bd93e6
101259 F20110330_AABRNA gezan_s_Page_019.jp2
6eaf3afbce344c0a6d3b68531e069a0f
2926e75a7b2efa43666f2e0c12b2bc1054ee5d03
16875 F20110330_AABRMM gezan_s_Page_111.QC.jpg
68c857eabf365818a8ea5f688fb9644e
b38fe0896f961750a5de088483b35a4a85ee01e9
88782 F20110330_AABRLY gezan_s_Page_101.jp2
ce847ceed4ed1cebb1bcd566c3194ca3
5c4ec9ed3bf0d108ebecbf4e3c43aeeaeefac9a6
1051958 F20110330_AABRNB gezan_s_Page_009.jp2
758bb17a668db01e8b8943ac1e884c34
1673ba46a39180b2b13b5081880011431bd3648a
1311 F20110330_AABRMN gezan_s_Page_106.txt
90f100eada8ff73a7c3e6f2244f10201
4beb98b88085c24e873ec3013d078d4ab9463e0f
6051 F20110330_AABRLZ gezan_s_Page_010thm.jpg
15c2596ea7d1a680c2b72e69821de3d4
3f58157e96fa67bfb34fa5ca188dde765f7248d9
55252 F20110330_AABRMO gezan_s_Page_051.pro
3e2f23142390dfdd6ee90a92f8ce2141
9a4ea7e41dfb5954742309790b95b4db8cf7c479
25031 F20110330_AABRNC gezan_s_Page_034.QC.jpg
7f4d40b1b13a0fdd61daf098cf0683e7
19b599e06710f8a4372c363d663bc45220e35701
43976 F20110330_AABRMP gezan_s_Page_012.jpg
32240fb1dbd1878edad37a2cd6aa68ef
80c3f098417e8d6cc505cbaeb1f19253da6a72f8
74379 F20110330_AABRND gezan_s_Page_071.jpg
8d27cb3fb32379dc8660f377fc85be99
125c09b025b684b411a7764fce09e2bd1ebd784e
73929 F20110330_AABRMQ gezan_s_Page_082.jpg
597bf1f7f27377c08dcae939514e9a83
3b92038d788adc99fb56f8cfc19bc0fe3c3e523b
2204 F20110330_AABRNE gezan_s_Page_123.txt
6257c8870b957e47ccaf8f8776e36f39
384356ee7520cba16b0283f1d037751ef64c1489
F20110330_AABRMR gezan_s_Page_121.tif
2b990eaa2b2fc74a44f0a3e6b7ca3ef6
a77faf9cc31cae2a42f51d9bf8d3c55b315c70bf
44916 F20110330_AABRNF gezan_s_Page_110.pro
e0df0f2098add5def348a972bf9032be
24db199b2c177522ab5981659db668dff1f0d488
1526 F20110330_AABRMS gezan_s_Page_025.txt
6e36fb72cbf2cba0f73bcbe6d178a09b
f8e89db6180cfeebc6ca5cd74edd1aa325bb81c7
48475 F20110330_AABRNG gezan_s_Page_048.pro
715479272976c8dd3fddc6870eac52f5
68e54aa611d67e083d4aa70c8c700e4e5bb44fd2
48385 F20110330_AABRMT gezan_s_Page_072.jpg
79badac393735720fc2eb3b081be23c4
d49ba66f1b5606c8e7afce339c6b8f26adecca8f
17939 F20110330_AABRNH gezan_s_Page_056.QC.jpg
e02bb6fc40769d9f63ada2b2eb864cb9
d9a4b0a065c21e25b9893d45ba8191cdd9e70366
1831 F20110330_AABRMU gezan_s_Page_126.txt
2cc451b432978f4dec4b8eebc9bd7043
6739f4e55079be30a84e054c96b447b0edeece40
39270 F20110330_AABRNI gezan_s_Page_014.pro
4f231a6d38187db179606d60964c0664
93c9e5000469b050f81ffecbdb0b787614d189c8
58417 F20110330_AABRMV gezan_s_Page_129.pro
4b9a2b7ce027d9d1680a252dca2eefeb
9fe6a36eed948033351b669bb6c87ebc8cd1810f
22645 F20110330_AABRNJ gezan_s_Page_070.QC.jpg
4367be448bce4d022af4fd3d3333eff9
fac19336bce4951f535269c6895eda6c41bba2ef
F20110330_AABRMW gezan_s_Page_090.tif
5bd5a4887cb372fbf5fd0abc88594285
9ac4febcb30aba57f96438a1bc47477c92cf0d4c
F20110330_AABRNK gezan_s_Page_076.tif
9de63d19022733da93a852534d6d4bd4
d8e060ee6cbee22a8a90e615ce344a5246c07b1e
24442 F20110330_AABRMX gezan_s_Page_073.QC.jpg
b8ca0e0d45cf5616f10b6fd4fce28b24
edb55a24f11aa0ce98fc3cbc7cf6f0e6df8a827e
5998 F20110330_AABROA gezan_s_Page_096thm.jpg
c375d1d8198982797819dbeba2e3cec4
d5e75d40fda13a173d84caed4ef181aa73de6cf0
24790 F20110330_AABRNL gezan_s_Page_027.QC.jpg
85d9d35070a39976e1459c23b2dbd452
1eea57735945c18b8ec20985a715c4b88b55277a
F20110330_AABROB gezan_s_Page_057.tif
0db5fbc4dbbb5d1f7011650a95f68382
35b3091ef49aa4eb4df9d719400c6dd6c850fdaf
71761 F20110330_AABRNM gezan_s_Page_009.pro
e410037f6825453affe16d9b47f04c4e
6b433cc1ef5ad8654980a7a478415a940b358023
F20110330_AABRMY gezan_s_Page_061thm.jpg
97aace05c9ce143412383391cc2a6e44
f30b15b0b0d5509c2d96bfddac0b84a41254d2e6
88757 F20110330_AABROC gezan_s_Page_091.jp2
885d5fbc7a01509d48e3984b3601593d
77601e1757eb8fb913b4982362c4dfc3f3869b4d
1454 F20110330_AABRNN gezan_s_Page_132.txt
0ed3a2124f96ec16ddcbcee3bd94d090
abb83fb8ce72bc38835168854a9af080de5a2947
101697 F20110330_AABRMZ gezan_s_Page_080.jp2
2400b063d88ce389b8a8f9633a2747f2
53c1ab23f2f7893a9d2c185c981bd9e3061cc9c8
F20110330_AABRNO gezan_s_Page_064.tif
b445c27f6fa69dbf4762d72836bae46a
8fff1719d4b2a35604f62628e3f4f17cf310ba17
22855 F20110330_AABROD gezan_s_Page_116.QC.jpg
13a7b0d14b531f0bf5286df488a34f45
3d084f68ee263ad39d88fc4125e5379508476cf6
1907 F20110330_AABRNP gezan_s_Page_097.txt
a8c668a454120388c419f139834a3185
0448c993676aaab3bc155a75ec220d0e2c6eab9a
5405 F20110330_AABROE gezan_s_Page_043thm.jpg
258c4da5ed649c1e46ae7737e2171ae7
3f2df0d1df61065d74fb129f543a1933df023fff
81588 F20110330_AABRNQ gezan_s_Page_065.jpg
8dee394516e6a5fedda3fe178cfa87af
4e0ff2ca7e6042ef23bfc3a459997be2df40233e
4814 F20110330_AABROF gezan_s_Page_006thm.jpg
66ccb701376e0771c006ecc1917facf4
f6689e1823d717184b2743e3131ecd0e94e11589
1940 F20110330_AABRNR gezan_s_Page_033.txt
a77597798e4561268272ef394fe272ff
7edfbe92187aa825a46e7ab39024eb8d21f8b9de
F20110330_AABROG gezan_s_Page_032.tif
82263b6167b99108212b761d3e4ffa94
bdbb44519b4a8af944fc86e64e511c7ce51d8f4e
1051843 F20110330_AABRNS gezan_s_Page_035.jp2
7cb5f0f156862aae8c21d789323c3f1c
c5928045b132ab13f9357d5a69b7a144b824f27e
483010 F20110330_AABROH gezan_s_Page_052.jp2
9e5c5122f43fae34df6cccfc089e3567
55191a1816ef252f8322b86c88559ef87d76aac6
6020 F20110330_AABRNT gezan_s_Page_121thm.jpg
27e1935a1f13bff927c9a0dc12f9ed79
3043f89140919c68aa75ac5abf45b4c2055d9520
1426 F20110330_AABROI gezan_s_Page_103.txt
799de852445dd9d5c178cd145932f44c
4a840516d3b5d4c624cadc1f14341949240bfab4
957767 F20110330_AABRNU gezan_s_Page_075.jp2
b57b0eee28500d29219eb38d86257da2
da0a2ebf00ae29c89ff89d86d7c324e1b2031ed8
F20110330_AABROJ gezan_s_Page_045.tif
ed8655009f7681ee428a26bed565373e
b6ea25ade549cdf97e9bada5ddcfa75a7e327308
6265 F20110330_AABRNV gezan_s_Page_024thm.jpg
b66a6e200c13e352831d6fedb7cba3d3
7fa62be3111bcf2080c6d93ad36a200ed5076c83
F20110330_AABROK gezan_s_Page_072.tif
2d9358efb3264b9872fbef012e846c30
078caa35ca327c2a22bcae152c1ad1a003e655db
80821 F20110330_AABRNW gezan_s_Page_069.jpg
be76ade4ae1b4ae09b5bae4626f5a3ce
740fa732708cc8dc964e2d76c1f5bef982a9d0bd
3354 F20110330_AABROL gezan_s_Page_006.txt
98f71f91e8af9053f586a580f3129506
a55a189de1bc453b5bb18a49d77ef1c5d5b7e35f
68021 F20110330_AABRNX gezan_s_Page_013.jpg
c01a9264476f5adbc25fed0c034b455d
30680b96407bd3a5790dd0b059e06b29f152136e
84526 F20110330_AABRPA gezan_s_Page_102.jpg
502aafd65a53c1f17c0ca41829290635
c0ca044c8da76f6492af63927803a8d02265bc4a
79955 F20110330_AABROM gezan_s_Page_046.jpg
2ba68a71ff5d4e2156bafca334e71aa5
71548c76e6a318dff8f324b3866dad934af908e4
F20110330_AABRNY gezan_s_Page_078.tif
b1cbf0a23beb2fcf4fc7480c0733a1b3
cc97c42ce9240b9dfce433853f0c4de5389e2315
49412 F20110330_AABRPB gezan_s_Page_027.pro
7846d377f65267890cc450904a43a046
d72343426d99fe0785f66467425bb7d490200dd4
65088 F20110330_AABRON gezan_s_Page_124.jp2
c1e308cbda48afa74fd2b4192155716a
a520d0c14d288fc041735162e37cdcbf973d2d08
42518 F20110330_AABRNZ gezan_s_Page_021.pro
4a43b61837ae8164a6ead1af3d2d718e
e6910158ab7a3875ed61d76aca8dcfe6312ac6ca
5822 F20110330_AABRPC gezan_s_Page_045thm.jpg
4e9c685bcebefd9daea444e648954e88
c4003e11f3fc774c2bc309786fdb3cfd2b4aaa2f
24109 F20110330_AABROO gezan_s_Page_060.pro
ef0701c83f18d40e67e757e2a4f69d0c
ef06e3a61526547a06780ad5c9c09bb1af5ecea0
F20110330_AABRPD gezan_s_Page_026.tif
a1974f42bdadde4c90e71e427a621ffe
de0393e60655937e709be3ce3706a9df80cc1c2a
742786 F20110330_AABROP gezan_s_Page_056.jp2
66c9f909e911817fa50146e57eae7a84
f8e30c1d28629774c2e98ecbd825cfd2e2be133c
6183 F20110330_AABROQ gezan_s_Page_066thm.jpg
ca42f2c29911634593b7d56927a5ff30
39dc5b2ba95ec5d861fff7e5f26f2596266533d3
81992 F20110330_AABRPE gezan_s_Page_044.jpg
d05074f96474d28ef41bbb1c40a18673
86825a62abfc5e72bf84ebc8efcdfd1925a81eb5
4702 F20110330_AABROR gezan_s_Page_026thm.jpg
e9090a3129b3cb125193945817bd4bae
4ce36507a0a9efbe5914a7b4151cd2795298c4ff
1492 F20110330_AABRPF gezan_s_Page_004.txt
c0f090b387dc96de32b8e479b2d64a58
2ce9bd92327ea1ddd66fe828db89934539f627e5
F20110330_AABROS gezan_s_Page_096.tif
d3c2279fb0d50ab0d59686ace06c9a3a
87a39012e6a063d8b39f5c004310cf907cbb41d4
103227 F20110330_AABRPG gezan_s_Page_097.jp2
b1eedb4177cf7b0bd5afed279b52656b
cef1874823cb6eaf1af6cd64a80d4ac29628b8d7
F20110330_AABROT gezan_s_Page_030.tif
0fb5bb710a6d58d704608afca76d2b24
5d658262c638673da87c91abf683d69566812a3f
114185 F20110330_AABRPH gezan_s_Page_009.jpg
db13c4736bba176efe7ca6c0e21905eb
91390dc529bd4ab10204aea9138aedfe1b1166c8
83911 F20110330_AABROU gezan_s_Page_086.jpg
98aaca09e874faded220595cf2c108f7
86125e20085da42daf9e9d531cc89567064464ed
107490 F20110330_AABRPI gezan_s_Page_044.jp2
b35407192622c993663f0a118734f11a
b60912d5d94a83e6020460817b2a9ed2aa5c4277
2010 F20110330_AABROV gezan_s_Page_050.txt
c2b77719d2a1f142d43499210b78f46b
3467e571fa476a4d8b03db302c33014458713a3e
F20110330_AABRPJ gezan_s_Page_021.tif
d1188db2a441f0cdddf04b7b00156735
90b9e80f5c335ea67c37e7c620a63488fd0e9af2
6447 F20110330_AABROW gezan_s_Page_001.QC.jpg
0cd4f25c76eb4b725fa34982f06b4ed1
b89931b7da98cfcc0997567e905998d7527e2709
2495 F20110330_AABRPK gezan_s_Page_026.txt
12824c5dfc0ddf364619398f945e7c3d
f536b8ec3ab0ef50814e860f5ab6bcea69485d61
50465 F20110330_AABROX gezan_s_Page_023.pro
2a5777035c24f4dfc87f5e9933c5241b
ee6832c2f78cfde6f06293b6745cded6073acafe
F20110330_AABRQA gezan_s_Page_046.tif
b96f2d819e6c4a88f499b10f72392214
a626c95b730bef7c973ed7bff24a68dfee325fed
79553 F20110330_AABRPL gezan_s_Page_029.jpg
cb91415c227d296402afef734ea84e7c
33e2eb41da0f1063a430d3a9a93bfc5803931462
30558 F20110330_AABROY gezan_s_Page_125.jpg
90de6193b1810e2b508fe0181a29c7d2
c639ea8f143081dffade455113d8f7450b150536
F20110330_AABRQB gezan_s_Page_071.tif
a231d71007a7925a8b630fa9f25a12c8
96c17917fb00f781835641ff0ffd7fee74f88a53
32228 F20110330_AABRPM gezan_s_Page_095.pro
30dbb8890b1dd47593555ba410686884
0986d741df264c043fce9f064c4152a6b36426c3
25901 F20110330_AABROZ gezan_s_Page_057.QC.jpg
a257dd25508e81c9c5636f3fcdd65b3c
2804ae660791ff947a7dcf70a4ccbeb6a246be55
51889 F20110330_AABRQC gezan_s_Page_049.pro
700bbafb4db3b4e5b61f9f1ea7cba262
41395d03496005eec88cac7709d9745102db8bd1
12432 F20110330_AABRPN gezan_s_Page_020.QC.jpg
2e05410a49c56215160a5f6e8a0b9a6b
e156f8433219c33c09cf5a038a44165f6e531cd6
63763 F20110330_AABRQD gezan_s_Page_121.jpg
137327bdc0cd09507fe8b44ac353b1d9
88cc17f7c3d4e4568e9e19d1ef1dece88912158f
105867 F20110330_AABRPO gezan_s_Page_127.jpg
f454da101c2d58bd3ae59b7e67287273
446a4f43c90b0c7104466ae7b7a38599a1b091ab



PAGE 1

OPTIMAL DESIGN AND ANALYS IS OF CLONAL FORESTRY TRIALS USING SIMULATED DATA By SALVADOR ALEJANDRO GEZAN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2005

PAGE 2

Copyright 2005 by Salvador Alejandro Gezan

PAGE 3

"If a scientific heresy is ignored or denounced by the general public, there is a chance it may be right. If a scientific heresy is emoti onally supported by the general public, it is almost certainly wrong." (Isaac Asimov, 1977) Dedicated to: My biological family, Tita, Lincoln, Ivan, Demain, Florencia and Alexandra My political family, Dean, Quena and Becky, Lauren, Anha and Kevin And to an special individual that put us all together, Pincho

PAGE 4

iv ACKNOWLEDGMENTS I would like to thank the me mbers of my supervisory committee, Drs. T.L. White, R.C. Littell, D.A. Huber, D. S. Wofford, a nd R. L. Wu, for their energy, time and help during my program. I thank, in particular to Dr. White, for the opportunity to come to Gainesville and do such an interesting projec t and also for his patience and wiliness to show me not only science, but also emotional intelligence. I thank to Dr. Huber, for many small but important details, and for show ing me a different di mension of thinking, sometimes unreachable but which surprisingly I liked. I thank to Dr. Littell, an unconditional supporter and also model in many senses. This research would not have been d one without the financial support of the members of the Cooperative Forest Genetic Re search Program (CFGRP). Here, I want to thank Greg Powell for his enormous support in several aspects, and for showing me what being a gator is all about. I also want to thank several friends: Th e Latino Mafia, Vero nica, Alex, Rodrigo, Bernardo, Belkys, Gabriela, and Rossanna; the Trigators Mafia, Terrence, Mathew, Shannon, Josh, Betsy, Mark, and with special love Eugenia. Special thanks go to all members of my family for tolerating my absence, but particularly to my step father Dean W. Pettit, for his unusual vision and for opening doors for me to cross as I choose.

PAGE 5

v TABLE OF CONTENTS page ACKNOWLEDGMENTS.................................................................................................iv LIST OF TABLES............................................................................................................vii LIST OF FIGURES.............................................................................................................x ABSTRACT.....................................................................................................................xi ii CHAPTER 1 INTRODUCTION........................................................................................................1 2 COMPARISON OF EXPERIMENTAL DESIGNS FOR CLONAL FORESTRY USING SIMULATED DATA......................................................................................7 Introduction................................................................................................................... 7 Materials and Methods...............................................................................................10 Field Layout.........................................................................................................10 Genetic Structure and Linear Model...................................................................12 Spatial Surface.....................................................................................................13 Simulation Process..............................................................................................15 Simulating Mortality...........................................................................................16 Statistical Analysis..............................................................................................17 Results and Discussion...............................................................................................19 Variance Components.........................................................................................19 Correlations between True a nd Predicted Clonal Values....................................22 Heritabilities for Simulated Designs...................................................................24 Conclusions.................................................................................................................26 3 ACHIEVING HIGHER HERITABILITI ES THROUGH IMPROVED DESIGN AND ANALYSIS OF CLONAL TRIALS.................................................................29 Introduction.................................................................................................................29 Materials and Methods...............................................................................................31 Field Design and Simulation...............................................................................31 Statistical Models................................................................................................33 Results and Discussion...............................................................................................36 Heritability Estimates an d Confidence Intervals.................................................36

PAGE 6

vi Surface Parameter Behavior................................................................................40 Latinization..........................................................................................................43 Number of Ramets per Clone..............................................................................44 Conclusions and Recommendations...........................................................................47 4 ACCOUNTING FOR SPATIAL VARI ABILITY IN BREEDING TRIALS...........49 Introduction.................................................................................................................49 Materials and Methods...............................................................................................53 Field Design and Simulation...............................................................................53 Statistical Analysis..............................................................................................55 Results........................................................................................................................ .59 Modeling Global and Local Trends.....................................................................59 Nearest Neighbor Models....................................................................................65 Discussion...................................................................................................................68 Modeling Global and Local Trends.....................................................................69 Nearest Neighbor Models....................................................................................72 Conclusions.................................................................................................................75 5 POST-HOC BLOCKING TO IMPROVE HERITABILITY AND BREEDING VALUE PREDICTION..............................................................................................77 Introduction.................................................................................................................77 Materials and Methods...............................................................................................79 Results........................................................................................................................ .83 Discussion...................................................................................................................88 6 CONCLUSIONS........................................................................................................93 APPENDIX A SUPPLEMENTAL TABLES.....................................................................................98 B EXTENSIONS IN CONFIDENCE INTERVALS FOR HE RITABILITY ESTIMATES............................................................................................................106 C MATLAB SOURCE CODE.....................................................................................108 Annotated Matlab Code to Generate Error Surfaces................................................108 Annotated Matlab Code to Generate Clonal Trials..................................................109 LIST OF REFERENCES.................................................................................................112 BIOGRAPHICAL SKETCH...........................................................................................118

PAGE 7

vii LIST OF TABLES Table page 2-1 Simulated designs and their number of re plicates, blocks per replicate, plots per block, and trees per block for single-tree plots (STP) and four-tree row plots (4tree). The number of rows and columns is given for the row-column design. All designs contained 2,048 trees arranged in a rectangular grid of 64 x 32 positions..12 2-2 Linear models fitted for simulated da tasets for single-tree plots (STP) and fourtree row plots (4-tree) over all surface patterns, plot a nd design types. All model effects were considered random...............................................................................18 3-1 Linear models fitted for non-Latinized experimental designs over all surface patterns. All model effects other than the mean were considered random..............33 3-2 Linear models fitted for Latinized li near experimental designs over all surface patterns. All model effects other than the mean were considered random..............33 3-3 Average individual single-site broadsense heritability with its standard deviations in parenthesis and relative efficiencies calculated over completely randomized design for experimental desi gns and linear models that were not Latinized (No-LAT). Note that the parametric HB 2 was established for CR design as 0.25...........................................................................................................37 3-4 Coefficient of variation (100 std / x) for the estimated variance component for clone and error on 3 surface patterns, differe nt number of ramets per clone and 3 selected experimental designs..................................................................................45 4-1 Linear models fitted for simulated da tasets using classical experimental designs to model global trends: complete randomized (CR), randomized complete block (RCB), incomplete block design with 32 blocks (IB), and row-column (R-C) design. All model effects other than the mean were considered random.................56 4-2 Linear models fitted for simulated da tasets using polynomial functions to model global trends: linear (Linear), quadratic (Red-Poly), and quadratic model with some interactions (Full-Poly). All variables are assumed fixed and the treatment effect is random........................................................................................................56

PAGE 8

viii 4-3 Average variance components for the surface pattern GRAD obtained from fitting the models in Tables 4-1 and 4-2 for the following error structures: a) independent errors (ID); and b) autoregressive without nugget (AR1 AR1). All values are the means from 1,000 simula tions and the true values were 2 g = 0.25 and 2 s + 2 ms = 0.75 for CR design.........................................................................62 4-4 Average variance components for the su rface pattern ALL obtained from fitting the models in Tables 4-1 and 4-2 fo r the following error structures: a) independent errors (ID); b) au toregressive without nugget (AR1 AR1); and c) autoregressive with nugget (AR1 AR1+ ). All values are the means from 1,000 simulations and the true values were 2 g = 0.25 and 2 s + 2 ms = 0.75 for CR design.................................................................................................................63 4-5 Average variance components for th e surface pattern PATCH obtained from fitting the models in Tables 4-1 and 4-2 for the following error structures: a) independent errors (ID); b) au toregressive without nugget (AR1 AR1); and c) autoregressive with nugget (AR1 AR1+ ). All values are the means from 1,000 simulations and the true values were 2 g = 0.25 and 2 s + 2 ms = 0.75 for CR design.................................................................................................................64 4-6 Average correlation between true an d predicted treatment effects (CORR) and average treatment variance component for iterated Papadakis methods PAP-6, PAP-8 and PAP-11 for the surface patterns ALL, PATCH and GRAD..................68 5-1 Linear models fitted for simulated da tasets using classical experimental designs to model global trends over all surfa ce patterns: randomized complete block (RCB), incomplete block design with x blocks (IB), and a row-column (R-C) design. All model effects other than the mean were considered random.................81 5-2 Average variance components and indi vidual broad-sense he ritability for all surface patterns obtained from fitting the models in Table 5-1 from simulated datasets with a randomized complete block designs with 8 ramets per clone per site........................................................................................................................... .85 5-3 Relative frequency of best post-hoc blocking fits from 500 datasets for each post-hoc design and surface pattern with 8 and 2 ramets per clone per site according to the log-likelihood (logL) and average standard deviation of the difference (SED). The values in bold co rrespond to the most frequent designs......87 A-1 Average variance component estimates for single-tree plot with no mortality in all surface patterns and experimental de signs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design...................................................................................................98

PAGE 9

ix A-2 Average variance component estimates fo r four-tree plot with no mortality in all surface patterns and experimental designs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design.................................................................................................................99 A-3 Average variance component estimates for single-tree plot with 25% mortality in all surface patterns an d experimental designs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design.................................................................................................100 A-4 Average variance component estimates fo r four-tree plot with 25% mortality in all surface patterns and experimental de signs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design.................................................................................................101 A-5 Average estimated correlations betwee n true and predicted clonal (CORR) with standard deviations for 0 and 25% morta lity obtained from 1,000 simulations for single-tree plots (STP) and four-tree plot row plots (4-tree) in all surface patterns and experimental designs.......................................................................................102 A-6 Average individual broad-sense heritabil ity calculated for single-tree plot (STP) and four-tree plot (4-tree) with their standard deviati ons in all surface patterns, experimental designs and mortality cases. All values are the means for 1,000 simulations and the base heritabi lity value was 0.25 for CR design......................103 A-7 Coefficient of variation (100 std / x) for broad-sense indi vidual heritability obtained from 1,000 simulations with 8 rame ts per clone for single-tree plot in all surface patterns and experimental designs........................................................104 A-8 Average correlations between true an d predicted treatment effects (CORR) with standard deviations in parenthesis in th e ALL surface patterns for 2 and 8 ramets per clone for selected experimental designs analyses and polynomial models fitted for the following error structures: independent errors (ID); autoregressive without nugget (AR1 AR1); and autoregressive with nugget (AR1 AR1+ )...105 A-9 Original and post-hoc blocking averag e individual broad-se nse heritabilities with standard deviations in parenthe sis for all surface patterns obtained from fitting the models in Table 5-1 from simulated datasets with a randomized complete block design with 8 ramets per clone per site.........................................105

PAGE 10

x LIST OF FIGURES Figure page 2-1 Replicate layout for randomized comp lete block design simulations for singletree plots (left) and four-tr ee row plots (right). The same 8 resolvable replicates were used for incomplete block and ro w-column designs. All simulations had 256 unrelated clones with 8 ramets per clone for a total of 2,048 trees...................11 2-2 Average proportions of restrict ed maximum likelihood (REML) variance component estimates (after correction so that all components sum to one) for single-tree and four-tree row plots for the no mortality case in all surface patterns and experimental designs simulated...........................................................21 2-3 Average estimated correlations betw een true and predicted clonal values (CORR) obtained from 1,000 simulations fo r single-tree and four-tree row plots in the 0% mortality case for all surface patterns and designs..................................23 2-4 Average heritabilities obtained for single-tree and four-tre e row plots in each surface pattern and design type for the no mortality case. The error bars correspond to the upper limit for a 95% c onfidence interval of the mean...............25 2-5 Heritability distributions for the no mo rtality case in the surface pattern ALL for selected completely randomized (CR), randomized complete block (RCB), and row-column (R-C) designs in sing le-tree and four-tree row plots...........................25 3-1 Plots of means and 95% confidence intervals for estimated individual broadsense heritabilities obtained from simu lation runs and by the method proposed by Dickerson (1969) for all design types and surface patterns for non-Latinized designs and analyses (No-LAT). Average heritabilities correspond to the central points in each confidence interval, and all simulati ons involved 8 ramets per clone.........................................................................................................................3 8 3-2 Contours for individual heritabilit ies obtained by using the lowess smoother (Cleveland 1979) of row-column design fo r different surface parameters from fitting experimental designs and analyses that were non-Latinized (No-LAT).......39 3-3 Average individual heritability estim ates for ALL and PATCH surface pattern for simulated surfaces for subsets of simulations with High ( x > 0.7 and y > 0.7) and Low ( x < 0.3 and y < 0.3) spatial correlations.........................................42

PAGE 11

xi 3-4 Distribution of indivi dual broad-sense heri tability estimates for row-column designs with 2 and 8 ramets per clone obtained from 500 simulations each for the surface patterns ALL and PATCH from fitting experimental designs and analyses that were non-Latinized (No-LAT)...........................................................46 3-5 Average values for row-columns design in all surface patterns from fitting experimental designs and analyses that were non-Latinized: a) Clonal mean heritabilities (the whiskers correspond to the 95% confidence intervals); and b) Clonal mean heritabi lity increments........................................................................46 4-1 Neighbor plots and definitions of co variates used in Papadakis (PAP) and Moving Average (MA) methods. Plots with the same numbers indicate a common covariate....................................................................................................58 4-2 Average correlations between true an d predicted treatment effects (CORR) in 3 different surface patterns for classical experimental design analyses and polynomial models fitted for the following error structures: independent errors (ID); autoregressive without nugget (AR1 AR1); and autoregressive with nugget (AR1 AR1+ ). Surface GRAD is not shown in the last graph because few simulations converged.......................................................................................61 4-3 Average correlations between true an d predicted treatment effects (CORR) in 3 different surface patterns for neares t neighbor analyses fitted assuming independent errors. PAP: Papa dakis, MA: Moving Average...................................67 4-4 Average correlations between true an d predicted treatment effects (CORR) in 3 different surface patterns for selected methods: randomized complete block (RCB), row-column (R-C), and Papadakis (PAP) using the following error structures: independent errors (ID ), and autoregressive with nugget (AR1 AR1+ )........................................................................................................67 5-1 Experimental layout for randomized comp lete block design simulations with 8 resolvable replicates (left) and partiti oning of one replicate for incomplete block design layouts (right). All simulations had 256 unrelated clones with 8 ramets per clone per site for a total of 2,048 trees...............................................................81 5-2 Average correlation between true and predicted clonal values (CORR) for original and post-hoc bloking analys es on ALL, PATCH and GRAD surface patterns for simulated surfaces with 8 ramets per clone..........................................84 5-3 Average individual heritability (a) an d correlation between true and predicted clonal values (b) calculated over 500 si mulations with a randomized complete block designs with 8 ramets per clone pe r site for each surf ace pattern and posthoc design.................................................................................................................86

PAGE 12

xii 5-4 Average correlation between true and predicted clonal values (CORR) calculated for ALL and PATCH surface patt erns and all post-hoc designs for sets with High ( x, > 0.7 and y > 0.7) and Low ( x < 0.3 and y < 0.3) spatial correlations. The datasets used corresponded to simulated surfaces with a randomized complete block design and 8 ramets per clone per site........................89 B-1 Plots of means and 95% confidence inte rvals for broad-sense heritabilities for single-tree plots (STP) and four-tree row pl ots (4-tree) in al l designs and surface patterns for non-Latinized designs and an alyses (No-LAT) with 25% mortality. The methods correspond to simulation runs and the approximation proposed by Dickerson (1969). Average heritabilities co rrespond to the central points in each confidence interval.................................................................................................107

PAGE 13

xiii Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy OPTIMAL DESIGN AND ANALYS IS OF CLONAL FORESTRY TRIALS USING SIMULATED DATA By Salvador Alejandro Gezan December 2005 Chair: Timothy L. White Cochair: Ramon Littell Major Department: Forest Resources and Conservation Various alternatives for the design and anal ysis of clonal field trials in forestry were studied to identify “optimal” or “near optimal” experimental de signs and statistical techniques for estimating genetic parameters through the use of simulated data for single site analysis. These simulations investigated the consequences of different plot types (single-tree or four-tree row), experimental designs, presence of mortality, patterns of environmental heterogeneity, and number of ramets per clone. Approximated confidence intervals, Latinization and post-hoc blocking we re also studied. Later, spatial techniques such as nearest neighbor met hods and modeling of the erro r structure by specifying an autoregressive covariance fitted with tw o variants (with and without nugget) were compared. Considerable improvements were obtain ed through selecti on of appropriate experimental designs and stat istical analyses. A 5% higher correlation between true and predicted clonal values was found for single-tree plots as compar ed to four-tree plots. The

PAGE 14

xiv best experimental designs were row-column for single-tree plots, a nd incomplete blocks with 32 blocks for four-tree row plots in creasing heritability by 10% and 14% over a randomized complete block design, respectivel y. Larger variability of some variance component estimates was the only effect of 25% mortality. Experiments with more ramets per clone yielded higher clonal mean heritabilities, and using between 4 and 6 ramets per clone per site is recommended. Dickerson’s approximate method for estim ating the variance of heritability estimates produced reasonable 95% confiden ce intervals, but an underestimation was detected in the upper confidence limit of co mplex designs. The effects of implementing Latinization were significant for increasing heritability, but sma ll in practical terms. Also, substantial improvements in statistical effi ciency were obtained using post-hoc blocking, with negligible differences compared to predesigned local control with no reduction in the genetic variance as the si ze of the block decreased. For spatial analyses, the incorporation of a separable autoregressive error structure with or without nugget yielded the best results. Differences between experimental designs were almost non-existent when an error struct ure was also modeled. Some variants of the Papadakis method were almost as good as models that incorporated the error structure. Further, an iterated Papadakis met hod did not produce improvements over the noniterated Papadakis.

PAGE 15

1 CHAPTER 1 INTRODUCTION Genetic testing is a critical activity fo r all breeding programs; however, it is timeconsuming and frequently one of the most expensive processes (Z obel and Talbert 1984, p. 232). Appropriate selection of experimental design and statistical analysis can yield considerable improvements in terms of increas ed precision of genetic parameter estimates and efficient use of resources. Genetic testing aims to achieve several objectives such as (White 2004) 1) improved genetic gain from se lection by evaluation of genetic quality; 2) estimation of genetic parameters; 3) creati on of a base population for future selection cycles; and 4) quantification of realized genetic gains. Als o, various operational decisions depend on information obtained from genetic tests, and limited quantitative genetic knowledge degrades the efficiency of a breed ing program, eventually affecting potential gains. At present, there is extensive research in genetic testing for agricultural crops that can be useful for testing in forestry; nevertheless, for fore st species several important differences must be considere d. Forest sites tend to be he terogeneous, the individuals of interest (trees) are usually large, and the screening of hundreds or thousands of genetic entities ( e.g. families or clones) is common. These el ements imply the need for relatively large test sites or the use of fewer replicat es per site for each entry. Finding optimal testing sites is difficult, and if areas with high environmenta l heterogeneity are selected, the residual error variance will be inflated due to the confounding of tree-to-tree variation with other large-scale environmental effects. Nevertheless, optimization of genetic testing

PAGE 16

2 is possible in any of the following stages: 1) design of experiments; 2) implementation; and 3) statistical analysis. The design of experiments is a critical element and no single design will suit all testing objectives and environmental c onditions perfectly, bu t statistical and computational tools are available to improve efficiency and they should be used as safeguards against site hete rogeneity or potential undetect ed biases. Traditionally, in forest tree improvement, randomized complete block designs have been the design of choice, which is effective when within replicat e (or block) variability is relatively small, a situation that is rare in forest sites (Costa e Silva et al. 2001). Nevertheless, more complex and efficient designs can be easily implemented. Incomplete block designs allow for a better control of site heterogeneity by specifying smaller compartments than do randomized complete block design. Also, row and column positions of the experimental units can be ut ilized to simultaneously impl ement two-way blocking factors producing row-column designs. Other design opti ons include the utiliz ation of restricted randomization such as Latinization, nested structures and spatial designs (Whitaker et al. 2002). In the second stage of optimizing experimental precision ( i.e. implementation), it is important to exercise control at all levels of field testi ng (installation, maintenance and measurement) with proper care on do cumentation, labeling, randomization, site uniformity and survival. The objectives of this stage are to control fo r all possible factors that might increase experimental noise, and, therefore, reduce precision of estimation of parameters of interest.

PAGE 17

3 The stage of statistical analysis is also very important. The use of mixed linear models combined with techniques such as restricted maximum likelihood (REML) to estimate variance components and to predic t random effects is well understood and broadly used. Spatia l analysis (Gilmour et al. 1997) and nearest neighbor methods (Vollmann et al. 1996) are interesting tec hniques used to improve prediction of genetic values. Spatial analyses are particularly attractive because th ey incorporate the coordinates of the experimental units (plots, trees or plants) in the linear model to account for physical proximity by mode ling the error structure ( i.e. environmental heterogeneity). Another relevant tool is posthoc blocking, which consists of superimposing complete or incomplete blocks over the original field de sign and fitting a modified linear model as if the blocking effects were pr esent in the original design. Clonal forestry is a new pract ice in many widely used comm ercial forestry species, e.g. Pinus spp., Eucalyptus spp. and Populus spp. (Carson 1986; Elridge et al. 1994, p. 230; Ritchie 1992). Clonal forestry refers to the use of a relatively small number of tested clones deployed in operational plantations through mass-propagati on techniques (Bonga and Park 2003). Testing clones is of particular interest because the use of identical genetic entities (ramets from a clone) allows sampling several environmental microsites, thus allowing complete separation of geneti c and environmental effects to increase the precision of prediction of breeding values (Shaw and Hood 1985). Other benefits of clonal testing include 1) ach ieving greater genetic gains than under traditional tree breeding; 2) capturing greater portions of non-additive genetic variation for deployment; 3) using genotype x environment interaction by selecting clones most suited to specific site conditions; 4) detecting and utilizing correlation breaker s for traits with undesirable

PAGE 18

4 genetic correlations; 5) preventing inbreeding in populations; 6) incr easing plantation and product uniformity; and 7) reducing time be tween breeding and testing cycles (Libby 1977; Libby and Rauter 1984; Zobel and Talbert 1984, p. 311). There is extensive research relative to fi eld testing of progenies using seedling from half or full-sib families ( e.g. White 2004). Some of these guidelines can be used for clonal testing, but differences must be taken into consideration. In the literature, few studies exist that give im portant guidelines for the design and analysis of clonal experiments. The characteristics of the optim al designs and analyses are influenced by many factors such as magnitudes of heritability, g x e interac tion, and ratio of additive to non-additive genetic variance. Also, several ques tions of interest rema in to be answered. For example, for the design of experiments: 1) are single-tree plots more efficient than multiple-tree plots? is there an optimal incomp lete block size? 2) how much better than randomized complete blocks are incomplete bl ocks or row-column designs? 3) what are the effects of mortality on the estimation of genetic parameters? 4) do the best designs depend on the pattern of environmental hetero geneity? and 5) what is the optimal number of ramets per clone? For statistical analysis we require answer s to the following: 1) is there any gain from incorporating the error structure in the m odel? 2) is it necessary to model gradients? 3) how good are the nearest neighbor methods compared with spatial analysis techniques? and 4) is posthoc blocking useful; and how should we implement post-hoc blocking? One of the best ways to answer many of the above questions is with the use of simulation techniques. Statistical simulation or Monte Carlo experiments are widely used

PAGE 19

5 in statistical resear ch. These methods are based on th e generation of random numbers with a computer to obtain approximate soluti ons to problems that are difficult to solve analytically, and they can aid understandi ng and knowledge of the properties of the experiments or methods of interest (Johnson 1987, p. 1). Particularly, for field testing, they allow several alternatives to be evalua ted without incurring la rge expense or having to wait years before practical results are available. The overall goal of this research was to identify “optimal” or “near optimal” experimental designs and statis tical analysis techniques for th e prediction of clonal values and estimation of genetic parameters in or der to achieve maximum genetic gains from clonal testing. This study explor ed single-site simulations with sets of unrelated clones “planted” in environments with di fferent patterns of variability. In the first two chapters many alternative designs for clonal experiments are compared. In Chapter 2, the consequences fo r the estimation of genetic parameters of different design alternatives were studied. Th e elements compared were 1) single-tree plot versus four-tree row plots; 2) several experimental designs; 3) no mortality versus 25% mortality; and 4) three different e nvironmental patterns. For Chapter 3, more detailed work was done with STP experiment s. Here, individual heritabilities estimates were examined in detail according to the different parameters used to simulate the environmental patterns. Also, confidence inte rvals for heritability obtained using the method proposed by Dickerson (1969) were compared with simulated percentile confidence sets; and the effects of using di fferent number of ramets per clone were investigated.

PAGE 20

6 Several statistical analysis techniques were studied in the last two chapters using single-tree plot experiments only. First, in Chapter 4, multiple statistical tools were used to account for spatial variability. The selected techniques included 1) modeling of global trends through the use of traditional expe rimental designs or polynomial models; 2) specification of a separable autoregressive error structure (with and without nugget); and 3) variants of nearest neighbor methods (Papadakis and Moving Average). Finally, Chapter 5 aimed to understand consequences of and to define stra tegies for post-hoc blocking. Some of the aspects evaluated in cluded comparing performance of original blocking versus post-hoc blocking for seve ral experimental designs; and defining strategies to select an optimal blocki ng structure within post-hoc blocking.

PAGE 21

7 CHAPTER 2 COMPARISON OF EXPERIMENTAL DESIGNS FOR CLONAL FORESTRY USING SIMULATED DATA Introduction Genetic field tests of forest trees are critical to tr ee improvement serving to estimate genetic parameters and evaluate provenances, families, and individuals. Genetic testing of forest trees is a time-consuming pr ocess and is frequently the most expensive activity of an improvement program (Zobel a nd Talbert 1984, p. 232). Therefore, it is of primary interest to maximize benefits by allocating resources efficiently (Namkoong 1979, p. 117). It is common in forestry trials to study the performan ce of a large number of genetic entries ( e.g. families or clones), and this implies the need for relatively large test sites or the use of fewer replicates per site for each entry. Because forest sites for field experiments are often inherently variab le, it is difficult to find optimal sites. The presence of this environmental heterogeneity inflates the residual variance due to the confounding of tree-to-tree variat ion with other, larger-scale environmental effects; and, hence, decreases the benefits of using simple experimental designs (Grondona et al. 1996). Within-site variability is caused by variation in natural factors such as soil, microclimate, topography, wind and aspect. In addition other forms of heterogeneity might originate from machinery, stock quali ty or planting techni que. Some conditions produced by topography or moisture might be easy to identify but, more commonly, environmental heterogeneity is only recognized after the fact as differential response to

PAGE 22

8 environmental conditions. Gradients across the site, local patches, and random microsite variance are the common types of variabilit y, and these sources may appear individually or in combination (Costa e Silva et al. 2001), with some sources more frequent in particular geographical areas. There are three stages during which optimi zation of genetic testing can occur: 1) experimental design and planning ; 2) implementation; and 3) statistical analysis. Forest genetic trials have traditionally employed ra ndomized complete block designs (RCB) or, more recently, incomplete block designs (IB). The RCB is most effective when the site is relatively uniform within replicates, which is ra rely the case in forest sites (Costa e Silva et al. 2001). Reducing the size of the unit by incorporating inco mplete blocks within a full replicate (IB designs) allows for better c ontrol of site heteroge neity because smaller blocks tend to be less variable than larger one s; therefore, IB designs have the potential to increase precision over RCB (Cochran and Cox 1957, p. 386; Williams et al. 2002, p. 120). For simulated forest genetic trials under several environmental conditions, Fu et al. (1998) report a 42% increase in the efficiency for IB over RCB. In a related study Fu et al. (1999a) found that most of the within-site variation can be controlled using 5 to 20 plots per incomplete block with better results in those cases where square blocks were employed instead of row or column blocks. Another experimental option is the ro w-column design (John and Williams 1995, p. 87). When the experimental units are located in two-dimensional arrays, it is possible to simultaneously implement two blocking factors (i nstead of one as in IB) corresponding to the rows and columns of the experiment. Gr eater efficiencies are expected when rowcolumn designs are used (Lin et al. 1993; Williams and John 1996). Qiao et al. (2000),

PAGE 23

9 using several wheat breeding trials, reported an increase of efficiency over RCB of 11% for row-column designs compared with 8% for IB. Clonal forestry is a new practice in many widely planted commercial tree species, e.g. Pinus spp. Eucalyptus spp. Populus spp (Carson 1986; Elridge et al. 1993, p. 230; Ritchie 1992). Interest in clonal forestry stems from its additional benefits such as ability to achieve greater genetic gains; potential to capture greater portions of non-additive genetic variation and genotype x environmen t interaction; increased plantation and product uniformity; and accelerated use of results from tree improvement by reducing breeding and testing cycles (Libby 1 977; Zobel and Talbert 1984, p. 311). In the agronomy and forestry literature there are several studies that compare design and analysis alternatives for breed ing experiments through real or simulated datasets. In forest genetics, the few simula tion studies available report some guidelines for future testing. Fu et al. (1998) showed that -designs were the most efficient arrangement for IB designs. For the major ity of the environmental conditions studied they were superior to an RCB design for the es timation of family means, but the benefits of IB over RCB were reduced as the leve l of missing observations increased (Fu et al. 1999b). In a related study for -designs, Fu et al. (1999a) reported that smaller incomplete blocks were more efficient for si gnificant patches when single-tree plots were established to estimate family and clonal means. These studies give important guidelines for the design of genetic experiments; nevertheless, several questions of interest remain to be answered: Are single-tree plots more efficient than multiple-tree plots? In relation to IB designs, how much better (or worse) are row-column designs? Is there an optimal incomplete block size? How is the

PAGE 24

10 phenotypic variance partitioned under different experimental designs? Also, it is common in previous simulation studies to screen a limited number of cl ones and to treat the genetic entries as fixed effect s in the linear model. In this study genetic entries were assumed to be random, allowing estimation of heritabilities and prediction of clonal genetic values for a range of simulated conditions. The present study is focu sed on identifying “optimal” or “near optimal” experimental designs for estimating genetic parameters and achieving maximum genetic gain from forestry clonal te sts. This study explores sets of unrelated clones tested on a single site through the use of simulated da ta created with different patterns of environmental variability. In particular, th e objectives of these simulations are to investigate the consequences for the estimati on of genetic parameters of 1) using singletree or four-tree row plots; 2) using comp letely randomized, randomized complete block, incomplete blocks of various sizes, or rowcolumn designs; 3) experiencing no mortality versus 25% mortality; and 4) planting on site s with different envir onmental patterns of surface variation (only patches, only grad ients, and both patches and gradients). Materials and Methods Field Layout All simulations were based on a single trial of 2,048 trees planted on a contiguous rectangular site of 64 rows and 32 columns w ith square spacing and 8 ramets for each of the 256 unrelated clones. The trees were “pla nted” in either singl e-tree plots (STP) or four-tree row plots (4-tree), corresponding to 8 and 2 plots pe r clone, respectively (Figure 2-1). A completely randomized design ( CR) in which clonal plots were randomly assigned to the field site was consider ed the baseline experimental design for each plot size. A randomized complete block design ( RCB), a variety of incomplete block designs

PAGE 25

11 (IB), and a row-column design (R-C) were al so implemented for bot h STP and 4-tree row plots (Table 2-1). There are many ways in which treatments (o r clones in this case) can be allocated to incomplete blocks. For the present work, -designs were used to obtain IB and R-C layouts. These designs can be obtained quickly and e fficiently for many design parameters (Williams et al. 1999). All IB and R-C designs were generated using the software CycDesigN (Whitaker et al. 2002) that incorporates an algorithm to generate optimal or near-optimal -designs. Outputs from 100 diffe rent independent runs were used for each design and plot type. CR a nd RCB layouts were generated using code programmed in MATLAB (MathWorks 2000). Figure 2-1. Replicate layout for randomized complete block design simulations for single-tree plots (left) and four-tree row plots (right). The same 8 resolvable replicates were used for incomplete block and row-column designs. All simulations had 256 unrelated clones with 8 ramets per clone for a total of 2,048 trees.

PAGE 26

12 Table 2-1. Simulated designs and their number of replicates, blocks per replicate, plots per block, and trees per block for sing le-tree plots (STP) and four-tree row plots (4-tree). The number of rows and columns is given for the row-column design. All designs contained 2,048 trees ar ranged in a rectangular grid of 64 x 32 positions. STP Design a Replicates Blocks / Rep b Plots / Block Trees / Block CR 1 1 2048 2048 RCB 8 1 256 256 IB 4 8 4 64 64 IB 8 8 8 32 32 IB 16 8 16 16 16 IB 32 8 32 8 8 R-C 8 32 rows 32 columns 4-tree Design Replicates Blocks / Rep Plots / Block Trees / Block CR 1 1 512 2048 RCB 2 1 256 1024 IB 4 2 4 64 256 IB 8 2 8 32 128 IB 16 2 16 16 64 IB 32 2 32 8 32 R-C 2 16 rows 16 columns a CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. b Rep, resolvable replicate. Genetic Structure and Linear Model The linear model used to generate the simulated data was ) ( ) ( ijk ms ijk s k ijkE E C y 2-1 where yijk is the response of th e tree located in the ith row and jth column of the kth clone, is a fixed population mean whic h was set equal to 10 units, Ck is the random genetic clonal effect, Es(ijk) is the surface error (or st ructured residual) and Ems(ijk) corresponds to the microsite random error (or unstructured residual).

PAGE 27

13 The total variance for the above linear model (Equation 2-1) considering all components as random is 2 T = 2 clone + 2 s + 2 ms. For simplicity, but without loss of generality, 2 T was fixed to 1. Further, the variance structure for all surfaces and designs was set to 2 clone = 0.25 and 2 s + 2 ms = 0.75; hence, the single-site biased broad-sense heritability HB 2 = 2 clone / ( 2 clone + 2 s + 2 ms) is 0.25 for completely randomized designs. Spatial Surface The spatial surface is a rectangular grid (x and y coordinates) composed of unstructured and structured residu als. The unstructured errors, Ems, correspond to white noise and can originate from measurement e rrors, planting technique, stock quality, and unstructured microsite variati on. The structured residuals, Es, are due to the underlying environmental surface, and were generated fr om two distinct patterns: gradients and patches. Gradients were modele d as a mean response vector t of size 2,048 x 1, at each of the positions of the 64 x 32 grid, empl oying the following polynomial function: ) ( ) (2 2 ci ci ci ci ci ci iy x y x y x t 2-2 where xci and yci correspond to the centered values xci = xi – x and yci = yi – y for the ith tree located in column x and row y; and and represent fixed weights on linear and quadratic components, respectively. Th is function defines a flat plane ( i.e. no gradients), when and are zero. Also, the primary environmen tal gradient was oriented along the short axis of the 64 x 32 rect angle to minimize variation w ithin a replicate, as an experiment would be laid out in the field. Patches were modeled by incorporating a covariance structure based on a firstorder separable autoregressive process (AR1 AR1), which is a variant of the exponential model used in spatial statistics (Littell et al. 1996, p. 305). This error structure has been

PAGE 28

14 previously used successfully for analyses of agricultural experiments (Cullis and Gleeson 1991; Zimmerman and Harville 1991; Grondona et al. 1996; Gilmour et al. 1997) and forestry trials (Costa e Silva et al. 2001; Dutkowsky et al. 2002). The AR1 AR1 error structure considers two perpendicular co rrelations, one for the x direction ( x) and the other for y ( y). The model defines an anisotropi c model, where the covariance (or correlation) between two observati ons is not only a function of their distance, but also of their direction (Cressie 1993, p. 62). The parameters x and y define the correlation between the structured residuals (Es) of nearby trees, and ther e is a positive relationship between the magnitude of these parameters and patch size. To generate the patches, a 2,048 x 2,048 variance-covariance matrix ( R ) was constructed and later used to simulate resi duals. The elements of this matrix were obtained as s ie Var2) ( for diagonal elements 2-3 hy y hx x s i ie e Cov 2 ') ( for off-diagonal elements 2-4 where hx = | xi xi’ |, hy = | yi yi’ |, i.e ., the absolute distance in row and column position respectively, between two tr ees. The other parameters were previously defined. This covariance structure also includes a nugget parameter ( 2 ms), which allows the modeling of discontinuities in covariance over very small distances. In real datasets the nugget effect reflects variabi lity that would be found if multiple measurements were made exactly on the same position or the mi crosite variation of positions very close together (Cressie 1993, p. 59; Young and Young 1998, p. 256).

PAGE 29

15 Simulation Process The first stage of the simulation consisted of the independent ge neration of surfaces over which different experimental designs were superimposed. Each surface or environmental pattern was produced by selec ting 5 parameters at random from uniform distributions (, x, y and 2 ms). The parameters were restricted to the following ranges: 0 to 0.05 for 0 to 0.0005 for 0.01 to 0.99 for the correlation parameters x and y, and 0.15 to 0.60 for 2 ms. Additionally, these correlations were restricted so their absolute difference was smaller than 0.85; and 2 s was calculated from 2 s = 0.75 – 2 ms. All these parameters were then used to generate the vector t and the residual variance-covariance matrix R The Cholesky decomposition of the R matrix was obtained and used together with t and an independent random vect or of standard normal numbers to obtain the correlated residua ls (Johnson 1987, p. 52-54), producing e ~ N( t R ). Later, a set of normal independent random residuals was incorporated to constitute the white noise ( i.e. Ems). After this, all components were a dded together and a standardization was performed to ensure that the total en vironmental variance was fixed at 0.75. The remaining variability (0.25) belongs to the clonal component of the linear model (Equation 2-1). Because all clones are unrelated, each set of 256 clonal values was generated as 500 independent standard normal vectors, which were scaled to have a variance of 0.25. Three surface patterns were implemented to generate 1,000 independent surfaces of each pattern: patches only (PATCH) with = = 0 in Equation 2-2; gradients only (GRAD) where x = y = 0 in Equation 2-4; and both patches and gradients together (ALL) with none of the parameters set to ze ro. Because of the standardization that was

PAGE 30

16 applied to all surfaces, the co mparison between different patterns must be viewed with caution. ALL includes both gradients and patc hes and so is more heavily “corrected” when the error variance was adjusted to 0.75; hence, the effect of each surface component is reduced when compared with PATCH or GRAD. The process of generating a simulated tr ial for a particular surface pattern and design consisted of selecting at random a surface and a vector of clonal values. Additionally, a random experimental design (lay out) was selected from those generated through CycDesigN or MATLAB code, and a simple partial randomization was performed that consisted of shuffling clone numbers. Finally, all components were added together to produce the response vector ( y ), which was stored with other relevant variables (x and y positions, replicate, block, etc.) for statistical analysis. Simulating Mortality Two levels of mortality were considered in the present study: 0% and 25%. The process of mortality was modeled using two independent components: clone and microsite. The clonal component assumed that some clones survived better than others because of resistance to diseas e, adaptability to site, or difference in stock quality. This component was generated as a random vector of standard normal numbers independent of their original genetic values ( i.e. no relationship with their original clonal values, Ck). For the microsite component, it was assumed that mortality occurs in clusters or patches; and it was simulated as a patchy surface with exactly the same error structure previously described using Equations 2-3 and 2-4 where 20% of the total vari ation corresponded to white noise or random mortality. The followi ng values were used to generate the mortality surfaces: = = 0, x = y = 0.75, 2 ms = 0.20 and 2 s = 0.80. Finally, the

PAGE 31

17 mortality process consisted of calculating an index for each tree based on a weighted average of both components (10% for clona l and 90% for the micr osite component), and according to this index the lowest 25% of the tr ees were eliminated. If all ramets (out of 8) were eliminated for any clone, then the process was repeated. The procedure previously described does not take into account changes in competition due to death of neighboring trees. This effect was not incorpor ated because it is expected to be more relevant several years after establishment and to affect some variables more than others. Statistical Analysis The data for each of the simulated trials were analyzed using ASREML (Gilmour et al. 2002) with the linear models specified in Table 2-2 (a ll effects were considered random). This software fits mixed linear models producing restricted maximum likelihood (REML) estimates of varian ce components and best linear unbiased predictions (BLUP) of the random effect (Patterson and Thompson 1971). Altogether there were 84,000 datasets analyzed (3 surface patterns x 7 experimental designs x 2 plot types x 2 levels of mortality x 1,000 simulations each). The output was compiled and summarized to obtain averages and standard deviations for heritability estimates a nd REML variance components for the response variable y In addition, empirical corr elations (CORR) were cal culated between the true and predicted clonal values. Single-site heritabilities for each simulation were calculated as e clone clone BH2 2 2 2ˆ ˆ ˆ ˆ for STP, and 2-5 e plot clone clone BH2 2 2 2 2ˆ ˆ ˆ ˆ ˆ for 4-tree. 2-6

PAGE 32

18 Because heritability estimates corres pond to a ratio of correlated variance component estimates, the simple average is not an unbiased estimator of the first moment. Hence, the following formula was used n n He clone clone B 2 2 2 2ˆ ˆ ˆ for STP, and 2-7 n n He plot clone clone B 2 2 2 2 2ˆ ˆ ˆ ˆ for 4-tree 2-8 where the sums are over n=1,000 simulations of each surface pattern, plot and design type. Table 2-2. Linear models fitted for simulated datasets for single-tree plots (STP) and four-tree row plots (4-tree) over all surface patterns, pl ot and design types. All model effects were considered random. Design a STP b CR yij = + clonei + ij RCB yij = + repi + clonej + ij IB yijk = + repi + block(rep)ij + clonek + ijk R-C yijkl = + repi + col(rep)ij + row(rep)ik + clonel + ijkl Design 4-tree CR yijk = + clonei + plotij + ijk RCB yijk = + repi + clonej + plotij + ijk IB yijkl = + repi + block(rep)ij + plotik + clonek + ijkl R-C yijklm = + repi + col(rep)ij + row(rep)ik + clonel + plotil + ijklm a CR, complete randomized; RCB, randomized complete block; IB, incomplete block; R-C, row-column. b clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within re plicate; row(rep), row nested with in replicate; plot, plot effect; residual.

PAGE 33

19 Genetic gain from clonal selection was not estimated in the present study because gain is directly proportional to heritability (HB 2) for a fixed selection intensity and phenotypic variance (Falconer and Mackay 199 6, p. 189). Thus, any increases in HB 2 lead directly to greater genetic gains. Also, the correlation between true and predicted clonal values (CORR) is a direct measure of genetic gain. As CORR approaches 1, clonal values are precisely predicted and genetic gain from clonal selection is maximized. Results and Discussion Variance Components For all surface patterns, designs and plot types the simulated data yielded REML estimates of the genetic variance that averag ed close to the parametric values imposed during simulation ( i.e. 2 clone = 0.25). Also, for CR (the refe rence experimental design) the estimated variance components for error aver aged almost exactly to their parametric values ( i.e. 2 = 0.75 for STP and 2 plot + 2 = 0.75 for 4-tree) (Figure 2-2). During the simulation process, the total variance was se t to 1, and, after analysis the estimated phenotypic variance (sum of average varian ce component estimates) were extremely close to 1 for CR for all surface patterns and plot types and for all experimental designs in PATCH (Tables A-1, A-2, A-3 and A-4). Simulation scenarios other than CR desi gns that contained gradients (ALL and GRAD) and replications resulted in estimated phenotypic variances sl ightly greater than 1 (2% to 6% overestimation) with higher valu es for 4-tree row plots than STP. This situation was due to an inflation of the erro r variance that originates when a replicate effect is incorporated in the model. Theoreti cally, a flat plane is specified for each level of replicate when in fact a non-horizontal trend should be used to model the field

PAGE 34

20 gradients. The presence of a slightly inflat ed total variance was corroborated in a small simulation study using surfaces that cont ained only gradient and no errors ( i.e. random noise or patches). Upon fitting an RCB to this gradient, residuals we re generated and an error variance was estimated when in fact it should not exist. Box and Hay (1953) reported a similar situation with time trends and indicated that this inflated error produced by trends within the replicates can be eliminated by using very small replicates. In this study, we preferred to correct the variance components in all models so that they will sum to 1 for each simulated dataset. For STP in PATCH and ALL surfaces, the variance component for the incomplete block effect ( 2 block(rep)) increased with the number of inco mplete blocks (Figure 2-2). For surfaces with patches. the error variance component, 2 decreased for smaller incomplete block sizes because an increasi ng portion of the variance was explained by incomplete blocks. In general, larg er values for the error component 2 were found for PATCH compared to GRAD surfaces indicating th at some of the variation due to patches tended to be confounded with the microsit e error instead of being captured by the incomplete blocks. Fu et al. (1998) reported a similar resu lt where blocks were more efficient in controlling for gradients th an for patches. The variance components 2 rep and 2 block(rep) for those surfaces with gradients (GRA D and ALL) explained a larger portion of the total variability than for PATCH surf aces. Also, for surfaces with gradients, the variance components for R-C indicated that there was more variability among columns than rows, since the main environmental gradie nt was oriented along th e short axis of the experiment.

PAGE 35

21 4-Tree row Surface: ALL CRRCBIB 4IB 8IB 16IB 32R-C 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 STP Surface: PATCH CRRCBIB 4IB 8IB 16IB 32R-C 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 STP Surface: ALL CRRCBIB 4IB 8IB 16IB 32R-C 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 STP Surface: GRAD CRRCBIB 4IB 8IB 16IB 32R-C 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 4-Tree row Surface: PATCH CRRCBIB 4IB 8IB 16IB 32R-C 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 4-Tree row Surface: GRAD CRRCBIB 4IB 8IB 16IB 32R-C 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Figure 2-2. Average proportions of rest ricted maximum likelihood (REML) variance component estimates (after correction so that all components sum to one) for single-tree and four-tree row plots for the no mortality case in all surface patterns and experimental designs simulated. Clone Rep Block(rep) Col(rep) Row(rep) Plot Error

PAGE 36

22 The tendencies for 4-tree were similar to STP except that the values of 2 rep were smaller because replicates were 4 times larger and more heterogeneous. Also, the proportion of variance explained by incomplete blocks was greater for 4-tree than STP for GRAD and ALL surfaces and sma ller for PATCH. But, the sum of 2 rep and 2 block(rep) was very similar between plot types with slig htly larger values for STP in PATCH. The plot effect ( 2 plot) was smaller for GRAD and larger for PATCH indicating that some of the variability of patches was confounded in th e plot effect. For both plot types, smaller values of 2 were found on ALL surfaces which had both gradients and patches. Similarly, Fu et al. (1998) found increased e fficiencies in the estimation of family means when high levels of gradients a nd patches occurred simultaneously. The variance components were very similar, comparing 25% mortality to 0% mortality, for all of the plot types, designs and surface patterns (Tables A-1, A-2, A-3 and A-4). The major impact of mortality wa s more variation in the variance component estimates. For 2 clone, the standard deviation increased from 6.8% to 12.6%, while for 2 the increase was only 0.2% to 2.5%. This re sult agrees with findings reported by Fu et al. (1999b) that showed for clonal tests a slight decrease in the efficiency of IB over RCB under random mortality. However, in the presen t study the impact of mortality was small considering that 25% of the observations were missing, and it demonstrates the usefulness of the REML technique to estimate unbiased parameters efficiently. Correlations between True and Predicted Clonal Values Performance of the various designs and pl ot types for predicting clonal values was assessed by the correlation between true a nd predicted clonal valu es (CORR). For all surface patterns the best 4-tree row plot design was less precise for predicting clonal

PAGE 37

23 values (lower CORR averages) than any STP design indicating that for these simulations the latter did a better job accounting for micr osite variation (Figur e 2-3). This finding agrees with results reported by Loo-Dinkins et al. (1990) and Costa e Silva et al. (2001). The difference in average CORR values between the best and worst experimental designs was larger for 4-tree than for STP, and th is difference increased in surfaces that incorporated gradients (GRAD and ALL). So, while 4-tree row plot designs were generally less efficient than STP, experi mental designs have more impact on the efficiency of experiments established with multiple-tree plots. For STP, the best designs were R-C fo llowed closely by IB 32. On GRAD surfaces all the incomplete block designs behaved similarly, and on PATCH the R-C performed best. For 4-tree the best desi gns for GRAD and ALL were IB 8, IB 16 and IB 32; and RC was clearly inferior, while for PATCH the best designs were IB 32 and R-C. 4-Tree row 0% Mortality CRRCBIB 4IB 8IB 16IB 32R-C CORR 0.74 0.76 0.78 0.80 0.82 0.84 0.86 0.88 0.90 STP 0% Mortality CRRCBIB 4IB 8IB 16IB 32R-C CORR 0.74 0.76 0.78 0.80 0.82 0.84 0.86 0.88 0.90 PATCH GRAD ALL PATCH GRAD ALL Figure 2-3. Average estimated correlations between true and predicted clonal values (CORR) obtained from 1,000 simulations for single-tree and four-tree row plots in the 0% mortality case fo r all surface patterns and designs.

PAGE 38

24 The effect of mortality on the average correlations was almost identical for all designs, plot types, and surface patterns. CORR decreased approximately 4% from the loss of 25% of the obser vations (Table A-5). Heritabilities for Simulated Designs As for variance components estimates, th e average heritability calculated using Equations 2-7 and 2-8 returned th e correct parametric value of HB 2 = 0.25 for the CR reference experimental design (see Figure 2-4). For the majority of all other experimental designs, the estimated average he ritabilities were always high er than 0.25 indicating that these designs were effective at reducing th e residual variance. For STP, there was a considerable increase in heritability for RCB above CR in GRAD and ALL surfaces (Table A-6 and Figure 2-4). For 4-tree row pl ots, the greater increase occurred when the design changed from RCB to incomplete bloc k designs. In STP, the design producing the highest average 2ˆBH for any surface pattern was R-C followed by IB 32. In the case of 4tree, the best design was clearly IB 32. For the IB designs, larger heritability values were obtained as the number of incomplete blocks increased for all plot type s and surface patterns i ndicating that smaller incomplete blocks were more efficient than larger blocks in controlling environmental variation. Similar conclusions were obta ined from other simulation studies where between 5 and 10 plots (or trees) per in complete block were recommended (Fu et al. 1999a). For this study, the best incomplete block design (IB 32) had only 8 plots per block for both STP and 4-tree row plots. This block size is rela tively small, and the use of fewer plots could produce grea ter heritabilities, but further studies are required.

PAGE 39

25 4-Tree row 0% Mortality CRRCBIB 4IB 8IB 16IB 32R-C Heritability 0.20 0.22 0.24 0.26 0.28 0.30 0.32 0.34 0.36 STP 0% Mortality CRRCBIB 4IB 8IB 16IB 32R-C Heritability 0.20 0.22 0.24 0.26 0.28 0.30 0.32 0.34 0.36 PATCH GRAD ALL ab Figure 2-4. Average heritabilities obtained for single-tree and four-t ree row plots in each surface pattern and design type for the no mortality case. The error bars correspond to the upper limit for a 95% confidence interval of the mean. 4-Tree row 0% MortalityHB 2 0.10.20.30.40.50.6 Frequency 0 50 100 150 200 250 300 CR RCB R-C STP 0% MortalityHB 2 0.10.20.30.40.50.6 Frequency 0 50 100 150 200 250 300 CR RCB R-C Figure 2-5. Heritability distributions for th e no mortality case in the surface pattern ALL for selected completely randomized ( CR), randomized complete block (RCB), and row-column (R-C) designs in si ngle-tree and four-tree row plots.

PAGE 40

26 The comparison of the distributions of HB 2 estimates for different designs showed that as the number of incomplete blocks increased the distribution became asymmetric, with a longer tail and increased spread to the right (Figure 2-5). This re sult indicates that assuming a normal distribution for heritability can sometimes be incorrect. A z-test comparing the 2ˆBH values for 0% and 25% mortality levels for designs of the same plot type and surface patterns wa s conducted to study if the heritability estimates changed. No statis tically significant differences were found (experiment-wise = 0.05) among average HB 2 estimates with the exception of a few comparisons within RC indicating that the analyses produced unbiased heritability estimates for all experimental designs and plot ty pes. The main effect of mortality was the increase in the variance among HB 2 estimates. Compared to 0% mortality, Var(2ˆBH ) increased 9.8% for STP and 12.0% for 4-tree row plots fo r datasets with 25% missing trees. The minor effect produced by 25% missing va lues could indicate that 8 ramets per clone was more than necessary. Several authors recommend using between 1 and 6 ramets/clone for single site experiment s (Shaw and Hood 1985, Russell and Libby 1986); hence, it is possible that under the conditions of the present study fewer than 8 ramets per clone might be adequate; nevertheless this topic requires further study. Conclusions The results from these simulations indicat e that proper selection of experimental designs can lead to considerable increases in heritability, precision of predicted genetic values and genetic gain from selection. The use of single-tree plots instead of 4-tree row plots resulted in an average increase in the correlation between true and predicted clonal values of 5%. Hence, single-tree plots allow a more effective sampling of the

PAGE 41

27 environmental variation and redu ce error variance more than 4-tree row plots. The best experimental design for singl e-tree plot experiments was the row-column design which fits random effects for both rows and column s within each resolvable replicate. R-C designs were followed closely by incomplete block designs w ith small blocks which were very efficient for STP. For 4-tree row plots, an incomplete block design with 32 blocks per replicate is recommended. The use of incomplete blocks (in one or two directions) controlled for an important portion of the total environmental variability and produced unbiased estimates of genetic variance components and clonal values. The in crease in the standard deviation of the average pair-wise clonal comparison from the use of incomplete blocks was counteracted by the improved precision obtained in the stat istical analysis. For single-tree plots, the smallest incomplete block under study had 8 trees per block, and it is possible that smaller blocks could produce even greater improvements. For the different simulated surface patte rns, the ranking using heritability or correlation from high to low was ALL, GRAD a nd PATCH. In the latter pattern some of the small patches were confounded with random error; hence, for this surface pattern lower heritability valu es should be expected. Twenty-five percent mortality produced only sl ight changes in the statistics studied. The consequences were primarily an increa se in the variability of some variance components (variance of 2 clone increases about 10%, and 2 about 1%) and, therefore, the variability of heritabilities increased. T hus, the effect of mortality was small, but might have been ameliorated by the relativel y large number of ramets used per clone.

PAGE 42

28 Lastly, the simulations from this study did not incorporate the effect of competition between neighboring trees; therefore, these results must be interpreted with caution. Trials with strong between tree competition, as occurs in older testing ages, and with large mortality patches, might produce different results.

PAGE 43

29 CHAPTER 3 ACHIEVING HIGHER HE RITABILITIES TH ROUGH IMPROVED DESIGN AND ANALYSIS OF CLONAL TRIALS Introduction For any operational breeding program, genetic testing constitutes one of the most important and expensive activities. Several al ternatives are available for experimental designs. The widely employed randomized complete block design (RCB) is most effective when blocks (or replicates) are rela tively uniform, something that usually occurs only with small replicates (Costa e Silva et al. 2001). With large numbers of treatments ( e.g., families and clones) the use of incomplete blocks (IB) can increase efficiency considerably (Fu et al. 1998; Fu et al. 1999a), particularly when there are large amounts of environmental variability or when the orie ntation of replicates can not be correctly specified (Lin et al. 1993). Also, the row and column positions of the experimental units can be utilized to simultaneously impl ement two blocking factors and produce rowcolumn designs (R-C), as described in detail by John and Williams (1995, p. 87). These designs have demonstrated greater effici encies than other common designs (Lin et al. 1993; Williams and John 1996). Latinization is rarely used as a design technique for increasing the efficiency of estimating trea tment effects. A design is said to be “Latinized” when the randomization is restricted so that the position of the experimental units for the same treatment are forced to samp le different areas of the experimental area. This is accomplished by defining long blocks (such as row or columns) that span multiple

PAGE 44

30 replicates ensuring that the treatments are spread out acr oss the entire te st site (John and Williams 1995, p. 87-88). The increased interest in clonal forestry is generated by the higher genetic gains that can be obtained from using tested clone s for deployment to operational plantations. Some benefits include the po ssibility of capturing non-addi tive genetic effects, using greater amounts of genotype x environment in teraction, and increasing plantation and product uniformity (Zobel a nd Talbert 1984, p. 311). Field experiments for clonal testing are particularly challenging because large numbers of genotypes need to be evaluated, implying the need fo r large test areas. Relatively homogenous areas are difficult to find because forest sites tend to have high environmental variability usually expressed in the form of patches, gradients or both, together with considerable random microsite noise (Costa e Silva et al. 2001). Therefore, site heterogeneity is an important factor to consider when clonal trials are designed, implemented and analyzed. Selection of an a ppropriate experimental design together with a correct specification of the linear mode l could produce considerable improvements in the precision of predicted genetic values, heritabilities and gains from selection. Several recommendations are available in the literature for clonal testing, and in general, optimal designs would employ 1 to 6 ramets per clone per site, and the number of clones tested would be maximized while ut ilizing as few ramets as possible (Shaw and Hood 1985; Russell and Libby 1986; Loo-Dinkins et al. 1990; Russell and Loo-Dinkins 1993). Still, difficulties remain in defining optimal designs, numbers of families, clones per family and ramets per clone to be plante d on single or multiple sites. Also, it is not clear which analytical met hods are to be preferred.

PAGE 45

31 In Chapter 2, we simulated single-site clonal trials under several experimental designs and 3 different environmental surface patterns to identify appropriate conditions for estimating genetic parameters. That st udy showed that single-tree plots (STP) experiments were more efficient for predic ting clonal values than four-tree row plots across all designs and conditions simulate d. This new study considers only STP experiments and aims to determine the effects of different designs and analytical options on the prediction of clonal va lues and the precision of so me genetic components. In particular, the objectives are to determine 1) which experimental designs, including Latinization, maximize broad-sens e heritability and, hence, gain from clonal selection? 2) which patterns of environmenta l or spatial variability yield high or low heritabilities? 3) what the effects of using different number of ramets per clone ar e? and 4) how close Dickerson’s approximate method for confidence intervals is to the empirical estimates? Materials and Methods Field Design and Simulation The simulated datasets used in this st udy are based on a single site trial of 2,048 ramets planted in a rectangular grid of 64 rows and 32 column s with 8 ramets for each of the 256 clones arranged in single-tree plots with no missing observations (details in Chapter 2). Briefly, two factors were cons idered in simulating the environments under which the different experiments were perfor med a gradient generated with a polynomial function depending on the x and y coordinates of the grid; and patches that were modeled by incorporating a covariance structure based in a first-order separable autoregressive process or AR1AR1 (Cressie 1993; Gilmour et al. 1997). The polynomial function employed was

PAGE 46

32 ) ( ) (2 2 ci ci ci ci ci ci iy x y x y x t 3-1 where xci and yci correspond to the centered values xci = xi –x and yci = yi –y for the ith tree located in column x and row y; and represent fixed weights on linear and quadratic components, respectively. The AR1AR1 error surface employed two perpendicular correlations ( x and y) and was generated through an error va riance-covariance matrix defined as Var (ei) = 2 s + 2 ms for diagonal elements 3-2 Cov (ei, ei’) = 2 s x hx y hy for off-diagonal elements 3-3 where hx = | xi xi’ |, hy = | yi yi’ |, i.e. the absolute distance between two trees in the row and column position, respectively. The variance component 2 s describes the surface error or structured residual, and 2 ms corresponds to the micros ite random error, white noise or unstructured residual. The three surf ace patterns simulated included only patches (PATCH), only gradients (GRAD), and a co mbination of both (ALL). Each surface was generated by drawing at random the parameters , x, y and 2 ms from uniform distributions with the follow ing ranges: 0 to 0.05 for 0 to 0.0005 for 0.01 to 0.99 for x and y, and 0.15 to 0.60 for 2 ms. Without loss of generality the total variance for all simulated datasets was fixed at 1, and the ge netic structure was set to have a single site biased broad-sense heritability of 0.25; hence, 2 clone = 0.25 and 2 s + 2 ms = 0.75. The experimental designs studied were completely randomized (CR), randomized complete block with 8 resolvable replicates (RCB), a variety of incomplete block designs with 4, 8, 16 and 32 incomplete blocks within each resolvable replicate (IB 4, IB 8, IB 16 and IB 32, respectively), and a row-column design (R-C). Both non-Latinized and Latinized designs were obtained for IB a nd R-C designs using the software CycDesigN

PAGE 47

33 (Whitaker et al. 2002). One thousand simulations were generated for each combination of surface pattern, design type and Latinizati on option, for a total of 36,000 datasets. Statistical Models All the simulated experiments were an alyzed using the software ASREML (Gilmour et al. 2002) fitting the models specified in Tables 3-1 and 3-2 for non-Latinized and Latinized design options, respectively. Fo r each set of simulations, the restricted maximum likelihood (REML) variance components estimates were obtained, and singlesite individual heritabilities for each f itted model were estimated using the following expression: e clone clone BH2 2 2 2ˆ ˆ ˆ ˆ 3-4 Table 3-1. Linear models fitted for non-Latinized experimental designs over all surface patterns. All model effects other than the mean were considered random. Design a Model b CR yij = + clonei + ij RCB yij = + repi + clonej + ij IB yijk = + repi + block(rep)ij + clonek + ijk R-C yijkl = + repi + col(rep)ij + row(rep)ik + clonel + ijkl a CR, complete randomized; RCB, randomized complete block; IB, incomplete block; R-C, row-column. b clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row(rep), row nested within replicate; residual. Table 3-2. Linear models fitted for Latinized linear experimental designs over all surface patterns. All model effects other than the mean were considered random. Design a Model b IB yijkl = + repi + latj + block(rep)ik + clonel + ijkl R-C yijklmn = + repi + lcolj + lrowk + col(rep)il + row(rep)im + clonen + ijklmn a IB, incomplete block; R-C, row-column. b clone, clone effect; rep, resolvable replicate; lat, Latinized block; block(rep), incomplete block nested within replicate; lcol, Latinized long column; lrow, Latinized long row; col(rep), short column nested within replicate; row(rep), shor t row nested within replicate; residual.

PAGE 48

34 To estimate an unbiased empirical expected heritability for each design and surface, the ratio of the mean of the numerator over the mean of the denominator was used (see Equation 3-4). Also, the relative efficiency (R E) was calculated as the ratio of means of the heritability of a particul ar design divided by that of an RCB design of the same surface pattern. In this study, Dickerson’s method was us ed to obtain approximate 95% confidence intervals (CI’s) on individual heritability es timates. Approximations are required because heritabilities correspond to a ra tio between linear functions of random variables, and in general, and particularly for unbalanced data, a closed form expressi on does not exist to estimate confidence intervals (Dieters 1994, p. 4). Assumptions and procedures suggested by Dickerson (1969) approximate the variance of 2ˆBH for each data set as 2 2 2 2 2ˆ ˆ ) ˆ ( ˆ ) ˆ ( ˆe clone clone Br a V H r a V 3-5 where the numerator ) ˆ ( ˆ2 cloner a V corresponded to the variance of a variance component obtained from ASREML based on asymptotic theory of restricted maximum likelihood estimation (Searle et al. 1992, p. 473). Then, Dickerson’s 95% confidence intervals for 2ˆBH were calculated as ) ˆ ( 96 1 ˆ2 2 B BH std H where ) ˆ (2 BH std is the square root of the variance expression from Equation 3-5. For co mparison, empirical or percentile CI’s were obtained by finding the top and bottom 2.5% estimated 2ˆBH from the 1,000 individual heritability estimates from each fitted set of simu lations. CI’s were calculated for each experimental design and environmental surface. The magnitude of 2ˆBH is affected by different levels of the parameters used to generate the error surfaces (, x, y and 2 ms). Influences of these parameters were

PAGE 49

35 studied using heritability estim ates obtained from the non-La tinized designs and models fitted without Latinization through the use of simple group means and by producing smoothed three-dimensional surfaces usi ng the loess smoother (Cleveland 1979). Two types of comparisons were made to st udy the efficiency of Latinization. First, the effectiveness of Latinization in the design phase alone was tested using a z-test with the estimated heritabilities that were generated with and without Latinization analyzed with linear models that did not in clude these Latinization effects ( i.e. only models from Table 3-1). These analyses correspond to Pa rtial-LAT (indicating La tinization in design but not in analysis) and No-L AT (no Latinization for both design and analysis). Second, to examine the importance of including Latini zation in the analysis phase, Latinized experimental designs were fitted with and without Latinization effects in their respective linear models. These analyses were identif ied as Full-LAT (indicating Latinization in design and analysis) and Partial-LAT, and the likelihood ratio test for mixed models described by Wolfinger (1996) was used to compare these fitted models. Finally, the effects of using different numbers of ramets per clone on the efficiency of estimating variance components and predictin g clonal values were studied by selecting subsets of each original simulated datase t and fitting their corresponding linear models. The first 500 simulations of three designs (RCB, IB 32 and R-C) previously generated for the non-Latinized designs were analyzed by sele cting at random 2, 4, 6 and 8 contiguous replicates and using the linear models fr om Table 3-1. Comparisons were preformed using both individual broad se nse heritability and mean clonal heritability ( 2ˆcH) from each subset. The 2ˆcH formula used was m He clone clone c/ ˆ ˆ ˆ ˆ2 2 2 2 3-6

PAGE 50

36 where the variance components are as previously described, and m is the number of ramets (2, 4, 6 or 8). As with Equation, 3-4 an unbiased mean was estimated as the ratio of the mean of the numerator ove r the mean of the denominator. In the present study comparisons of me thods and designs were made using heritabilities only. Genetic gains are also of interest, but they were not used here because for a given intensity of selection and fi xed phenotypic variance, gains are directly proportional to heritability (Fal coner and Mackay 1996, p. 189). Results and Discussion Heritability Estimates and Confidence Intervals The R-C design produced the highest averag e individual heritabi lity and relative efficiency (RE) for all surface patterns (Table 3-3 and Figure 2-4). Th is was particularly true for surface pattern ALL where both gradie nts and patches were simulated together. For IB designs, heritability increased as the size of the incomplete block decreased for all surfaces, and IB designs with 32 incomplete bloc ks per replicate (8 trees per block) were nearly as efficient as R-C designs for all surfaces. In general, for all designs higher relative efficiencies were found for surfaces with simula ted patches (ALL and PATCH), indicating the usefulne ss of one or two-way blocking to account for them. These results agree with other studies that also reported greater efficiencies fo r IB and R-C designs (Lin et al. 1993; Fu et al. 1999a; Qiao et al. 2000). Also, higher average heritabilities were accompanied by an increase in their sta ndard deviation (Table 3-3 and Figure 2-5). For example, in PATCH surfaces for R-C desi gns, the standard deviation was more than double that of RCB designs. Hence, the implem entation of genetic experiments with R-C and IB 32 designs is recomme nded; nevertheless, they have the drawback of yielding more variable heritability estimates.

PAGE 51

37 Table 3-3. Average individual single-site broad-sense heritability with its standard deviations in parenthesis and relative efficiencies calculated over completely randomized design for experimental desi gns and linear models that were not Latinized (No-LAT). Note that the parametric HB 2 was established for CR design as 0.25. 2ˆBH RE Design a ALL b GRAD PATCH ALL GRAD PATCH CR 0.250 (0.026) 0.252 (0.025) 0.251 (0.024) 0.84 0.86 0.98 RCB 0.296 (0.033) 0.291 (0.030) 0.256 (0.026) 1.00 1.00 1.00 IB 4 0.309 (0.040) 0.305 (0.033) 0.262 (0.030) 1.04 1.05 1.02 IB 8 0.317 (0.045) 0.307 (0.034) 0.269 (0.034) 1.07 1.05 1.05 IB 16 0.323 (0.049) 0.308 (0.035) 0.273 (0.038) 1.09 1.06 1.07 IB 32 0.328 (0.055) 0.308 (0.035) 0.284 (0.044) 1.11 1.06 1.11 R-C 0.336 (0.060) 0.312 (0.035) 0.287 (0.052) 1.13 1.07 1.12 a CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. b ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. In most cases, Dickerson’s approximation a nd the percentile CI’s obtained directly from simulations gave similar estimates for the 95% confidence intervals (Figure 3-1). Both methods were particularly close for th e RCB designs for all surface patterns, and for all designs in the GRAD surf aces. In contrast, Dickerson’s approximation almost always yielded smaller CI’s than the percentile in tervals for surfaces PATCH and ALL, and the upper limit of the CI’s was particularly unde restimated for R-C and IB designs on these surfaces. This was a consequence of Dicker son’s approximation assuming a symmetrical distribution when, in fact, the simulated 2ˆBH distributions were sk ewed to the right (see Figures 2-5 and 3-4).

PAGE 52

38 Surface: PATCH CRRCBIB 4IB 8IB 16IB 32R-C HERITABILITY 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 Surface: ALL CRRCBIB 4IB 8IB 16IB 32R-C HERITABILITY 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 Surface: GRAD CRRCBIB 4IB 8IB 16IB 32R-C HERITABILITY 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 Simulation Dickerson's Figure 3-1. Plots of means a nd 95% confidence intervals fo r estimated individual broadsense heritabilities obtained from simu lation runs and by the method proposed by Dickerson (1969) for all design types and surface patterns for nonLatinized designs and anal yses (No-LAT). Average he ritabilities correspond to the central points in each confidence interval, and all simulations involved 8 ramets per clone.

PAGE 53

39 H 2 Surface: GRAD 0.000.010.020.030.040.05 0.2 0.3 0.4 0.5 0.6 0.7 0.8 H 2 Surface: PATCH 0.00.20.40.60.81.0 0.0 0.2 0.4 0.6 0.8 1.0 H 2 Surface: PATCH 0.00.20.40.60.81.0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 H 2 Surface: GRAD 0.000.010.020.030.040.05 0.0000 0.0001 0.0002 0.0003 0.0004 0.0005 H 2 Surface: ALL x 0.00.20.40.60.81.0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 H 2 Surface: ALL x 0.00.20.40.60.81.0 y 0.0 0.2 0.4 0.6 0.8 1.0 2 0 2 0 2 0 x y xa b c d e f 0.25 0.30 0.35 0.40 0.45 0.50 Figure 3-2. Contours for indi vidual heritabilities obtaine d by using the lowess smoother (Cleveland 1979) of row-column design fo r different surface parameters from fitting experimental designs and analyses that were non-Latinized (No-LAT). 2 ms 2 ms 2 ms

PAGE 54

40 The ideal situation would be to derive the exact probability distribution for heritability, but this is very complex and us ually only possible for simple linear models and balanced datasets. Better approximate c onfidence intervals than Dickerson’s can be obtained using Taylor series, which improves the estimation of the heritability variance (more details in Dieters 1994, p. 17-20); neve rtheless, methods base d on Taylor series do not correct the problem of skewness. Anothe r option is to use boot strap and jackknife techniques to approximate distributions by re-sampling. These methods have been implemented successfully for specific situa tions with heritabi lity (Ndlovu 1992) and genetic correlations estimates (Liu et al. 1997), but their statistical properties are not well understood. Surface Parameter Behavior The simulation parameters (, x, y and 2 ms) used to generate the three surface patterns varied considerably and differentially influenced the magnitude of the individual heritability estimates. Gradients (or trends ) on simulated surfaces were specified by and here, larger values corresponded to steeper gradients. These parameters, in most cases, had little impact on heritability estimates; with larger values of and only small increases in heritabilities were found. Also, as expected, the effect of these gradient parameters was more pronounced in GRAD surf aces (for R-C see Figures 3-2c and 3-2d). It is well known that sm aller amounts of unstructu red residual increase the efficiency of analyses of fiel d experiments. For this study, an increase in heritabilities for all designs and surface patterns was noted as the simulated 2 ms decreased. This increase was more pronounced for designs with sm aller incomplete blocks, where some

PAGE 55

41 heritability estimates reached values as high as 0.50. Also, increases in 2ˆBH as the 2 ms got smaller were more pronounced in su rfaces with patche s (ALL and PATCH). For the simulated surfaces, using larger spatial correlation parameters ( x and y) produced bigger patches. The results of this study indicated that, in general, for all experimental designs there was an improve ment in heritabilities as the patch size increased. The effect of the spatial correla tion parameters was larger for IB and R-C designs than for RCB designs. Also, an interaction between these parameters and 2 ms was observed. For example, in R-C designs, hi gh values of spatial correlation and low values of white noise yielded la rger heritabilities, where the ef fect of spatial correlations was more pronounced at low values of 2 ms (Figures 3-2b and 3-2f). In another simulation study Fu et al. (1998) reported that increasing patch size in RCB and IB designs gave smaller variance of fa mily mean contrasts, a result that agrees with our findings. Nevertheless, they also reported that the presence of gradients produced a reduction in preci sion, which was more pronounced for RCB designs. In the present study, the effect of gradient was al most negligible, but more relevant for those designs that had smaller incomplete blocks The disparity of results is probably a consequence of using di fferent block sizes. Fu et al. (1998) considered a long rectangular block (10 x 1 tress) perpendicula r to the main gradient; therefore, increasing its effect, but in the present study blocks were always considered square or almost square. The larger effect of x and y on IB and R-C designs in comparison to RCB occurs because small incomplete blocks were more lik ely to be completely contained in a larger patch; therefore, reducing th e within block variab ility and explaining a larger portion of the total variance in comparison to using only la rge blocks or replicates. To examine this

PAGE 56

42 fact, the average individual heritabilities for subsets of simulations with low spatial correlations ( x < 0.3 and y < 0.3) and large correlations ( x, > 0.7 and y > 0.7) were compared. For all surface pa tterns, within the set of low correlations, no important differences were noted between all designs, but in the case of high correlations, an important increase in average he ritability occurred as the size of the incomplete block decreased. For example, in PATCH surfaces, the average heritability between IB 4 and IB 32 was fixed at 0.25 for low correlations, in contrast with a change from 0.29 to 0.35 for the high correlations set (Figure 3-3). T hus, the use of smaller incomplete blocks should be preferred for field experiments under the presence of modera te to strong spatial correlations. CRRCBIB 4IB 8IB 16IB 32R-C HERITABILITY 0.24 0.26 0.28 0.30 0.32 0.34 0.36 0.38 0.40 PATCH High PATCH Low ALL High ALL Low Figure 3-3. Average individual heritability estimates for ALL and PATCH surface pattern for simulated surfaces for subset s of simulations with High ( x > 0.7 and y > 0.7) and Low ( x < 0.3 and y < 0.3) spatial correlations.

PAGE 57

43 Latinization A z-test was conducted to compare average 2ˆBH values for all designs and surface patterns generated from simulations with and without Latinization but always fitted without the Latinizat ion random effects ( i.e. Partial-LAT versus No-LAT). This test indicated significant differences (experiment-wise = 0.05) for all IB and R-C designs on PATCH surfaces only. In this surface pattern, Latinized designs produced an increase up to 0.04 units of heritability, which tended to be larger for simpler designs. It was expected that the incorporation of a restrict ed randomization, as occurs with Latinized designs, would yield smalle r standard deviation for 2ˆBH and for the variance component 2 block(rep) (Williams et al 1999); but when Partial-LAT was compared to No-LAT very small reductions in heritability standard devi ations were found and changes in coefficient of variation were not relevant (for 2ˆBH see Table A-7). The other comparison considered only simulated datasets originating from Latinized experiments, which were fitted using linear models with and without Latinization effects ( i.e. Full-LAT and Partial-LAT). Th e results from the likelihood ratio test indicated statistical signi ficant differences (experiment-wise = 0.05) between these two models for all desi gns and surface patterns with the exception of designs IB 4 and IB 8 for PATCH surfaces. Better m odel fitting was found on Full-LAT, but differences in terms of average heritability, clonal and error varian ce were very small and not of practical relevance. Also, the la rgest improvements of including Latinization effects occurred in GRAD surfaces or with IB 32 designs in any surface. The above findings agree with what Qiao et al. (2000) reported for the anal ysis of several Latinized R-C designs, where greater efficiencies were re ported in those analys es that included the

PAGE 58

44 Latinized effects in the linear model. In summary, once a design is generated with Latinization, the incorpora tion of its corresponding random effects is important. The small benefits of Latinization found in this study in the design and analysis phase were not important in practice, but c ould possible be a cons equence of the large sample size used per clone (8 ramets). Nevert heless, the use of Latinization in the design stage is recommended to help control poten tial undetected environmental biases. Number of Ramets per Clone The variance components estimated thr ough REML for all experimental designs and surface patterns were very similar for data sets with different number of ramets per clone. Hence, even with reduced amounts of information, REML yielded unbiased variance component estimates and best linear unbiased predic tions of genetic values. The main impact of larger experiments ( i.e. more ramets per clone) was a decrease in variability or improved precisi on on variance component esti mates. The greatest overall reduction occurred when the number of ramets changed from 2 to 4, and was more pronounced for PATCH surfaces. For example, within R-C designs the coefficient of variation for the clonal variance component on average in all surfaces was reduced from 22.5% for analyses with 2 ramets to 12.0% for the case with 8 ramets (Table 3-4). As with variance components, for the same design type and surface pattern average heritability estimates changed very little be tween datasets with different numbers of ramets. Nevertheless, 2ˆBH distributions for any design type and surface pattern showed a decrease in range (and varian ce) as the number of ramets increased (for R-C design see Figure 3-4).

PAGE 59

45 Table 3-4. Coefficient of variation (100 std /x) for the estimated variance component for clone and error on 3 surface patterns, diffe rent number of ramets per clone and 3 selected experimental designs. Clonal Variance ramets / clone Design a Surface b 2 4 6 8 RCB ALL 21.5 % 14.4 % 13.1 % 12.0 % GRAD 23.6 % 14.6 % 12.7 % 11.7 % PATCH 24.5 % 15.4 % 12.7 % 11.8 % IB 32 ALL 21.5 % 14 .5 % 12.8 % 11.7 % GRAD 21.3 % 14.4 % 11.9 % 11.3 % PATCH 25.3 % 15.7 % 13.6 % 12.7 % R-C ALL 21.5 % 15.3 % 13.0 % 12.1 % GRAD 21.2 % 14.3 % 12.6 % 11.9 % PATCH 24.8 % 15.0 % 13.2 % 12.0 % Error Variance ramets / clone Design Surface 2 4 6 8 RCB ALL 15.0% 12.2% 11.0% 9.8% GRAD 15.9% 11.1% 9.9% 8.9% PATCH 10.5% 7.1% 5.0% 4.6% IB 32 ALL 22.0% 20.3% 20.0% 19.6% GRAD 14.9% 12.6% 12.3% 12.2% PATCH 17.1% 15.3% 15.0% 14.7% R-C ALL 22.5% 21.3% 20.8% 20.5% GRAD 14.1% 11.9% 11.6% 11.6% PATCH 18.9% 17.8% 17.3% 17.2% a RCB, randomized complete block; IB 32, incomplete block with 32 blocks per replicate; R-C, rowcolumn. b ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches.

PAGE 60

46 Surface: ALLHB 2 0.00.10.20.30.40.50.60.7 Frequency 0 20 40 60 80 100 120 140 160 2 8 Surface: PATCHHB 2 0.00.10.20.30.40.50.60.7 Frequency 0 20 40 60 80 100 120 140 160 2 8 Figure 3-4. Distribution of individual broad-sense he ritability estimates (2ˆBH ) for rowcolumn designs with 2 and 8 ramets pe r clone obtained from 500 simulations each for the surface patterns ALL and PATCH from fitting experimental designs and analyses that were non-Latinized (No-LAT). a ramets / clone 2468 Clonal Mean Heritability 0.4 0.5 0.6 0.7 0.8 0.9 ALL GRAD PATCH b ramets / clone Clonal Mean Heritability Increase 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20 ALL GRAD PATCH 2-4 4-6 6-8 Figure 3-5. Average values for row-column s design in all surface patterns from fitting experimental designs and analyses that were non-Latinized: a) Clonal mean heritabilities (the whiskers correspond to the 95% confidence intervals); and b) Clonal mean heritability increments.

PAGE 61

47 For all studied designs, as the number of ramets increased considerably higher clonal mean heritabilities ( 2ˆcH) were found; nevertheless, incremental improvements above 6 ramets per clone were small. For R-C designs 2ˆcHaveraged over all surfaces for 2, 4, 6 and 8 ramets per clone were 0.44, 0. 62, 0.70 and 0.76, respectively (Figure 3-5a); and the largest change occurred when the number of ramets changed from 2 to 4. Notably, these increments were always larg er in PATCH surfaces than any other surface (see Figure 3-5b). Hence, the results from these simulations (that considered no missing values) indicated that the optimal number of ramets pe r clone per site should be between 4 and 6. This number of ramets agrees with results obtained in other st udies that recommend between 1 and 6 ramets per clone per site (Shaw and Hood 1985; Russell and Libby 1986), and will produce near maximal genetic gains for clonal selection and will allow alternative uses of availabl e resources (for example, increasing the number of clones tested). Conclusions and Recommendations Appropriate selection of experiment al designs can yield considerably improvements in clonal testing. Row-column and incomplete block designs with a block size of 8 trees had the higher individual her itability estimates. Unfortunately, these two designs yielded more variable heritab ility estimates than simpler designs. Approximate 95% confidence intervals on heritability estimates obtained using Dickerson’s method were similar to empiri cal percentile CI’s. These methods were particularly close for simple designs (CR and RCB), but Dickerson’s approximation presented a bias in the upper confidence lim it of more complex designs (IB and R-C).

PAGE 62

48 The examination of the simulation parameters used to generate th e surfaces patterns had larger individual herita bilities when the amount of white noise or unstructured residual 2 ms was smaller and the spatial correlations larger, with a greater effect of the latter at low levels of 2 ms. Also, the effect of different levels of gradients on 2ˆBH tended to be almost non-existent, affec ting slightly GRAD surfaces only. In the present study, the use of Latinized designs improved the experimental design efficiency mainly for PATCH surfaces, but th e benefits of designing an experiment with Latinization were not important in practical te rms. On the other hand, once an experiment is designed with Latin ization, the inclusion of its co rresponding effects in the linear model yielded better analyses. As expected, experiments with more ramets per clone produced more precise variance component estimates a nd larger clonal mean heritab ilities. Using 4 to 6 ramets per clone per site is recomme nded. More than 6 ramets produced marginal improvements in precision of clonal means, but for clonal forestry it is important to screen large numbers of clones in one test. Finally, it is important to not e, that these findings were produced analyzing datasets that did not have missing values Nevertheless, in other rela ted studies (see Chapter 2 and Fu et al. 1999b) the effects of mortality were sm all. Therefore, these results can be extended safely to situations with low mortality levels.

PAGE 63

49 CHAPTER 4 ACCOUNTING FOR SPATIAL VARIABILITY IN BREEDING TRIALS Introduction The amount and type of environmental heterogeneity found in agricultural or forestry trials greatly influences the statis tical precision obtained in the comparison of treatments. As the homogeneity decreases, th e error variance of the treatment effect estimates increases and, consequently, it is ha rder to detect differences among treatments. Many natural and anthropogenic factors aff ecting site variability, such as soil, topography, wind, machinery or planting technique are usua lly detected after test establishment, making it difficult to implement optimal designs and to minimize the portion of the experimental erro r due to site variation (Fu et al. 1999). This environmental or spatial variability usually is expressed as gradients, patches or a combination of both, together with rando m microsite differences (Costa e Silva et al. 2001). Proper treatment randomization is enough to ensure that, over repeated testing and sampling, unbiased estimates of treatment e ffects are obtained, a nd hypothesis testing is valid (Grondona and Cr essie 1991; Brownie et al. 1993). Nevertheless, under environmental heterogeneity, the estimates obt ained for the analysis of a particular dataset are highly variable. This situa tion can be improved by the use of proper experimental designs and stat istical analysis. Randomized complete block, incomplete block and row-column designs constitute cla ssical approaches that control for spatial variability and are recognized as “ a priori ” techniques because they are implemented in

PAGE 64

50 the design stage. Another group of techniques (“ a posteriori” ) deals with statistical analyses that incorporate the x and y coordinates of the experi mental units (plots, trees or plants) in the linear model to account for physical proximity. The majority of statistical analyses for spatial data are based on modeling two main components of spatial variability (Grondona et al. 1996): 1) global tre nds or large-scale variation; and 2) local trends or small-scale variation. The first component corresponds to field gradients which can be modeled by: i) incorporating designs effects ( e.g. replicate, block, row, column, etc.); ii) using x and y positions as one or multiple fixed effects in the linear model to describe continuous ma thematical functions; or, iii) calculating covariates to adjust each observation. So me of the most common options are 1) polynomial regression (Brownie et al. 1993); 2) cubic smoothing spline (Durban et al. 1999; Verbyla et al. 1999); 3) nearest neighbor an alysis (Bartlett 1978; Brownie et al. 1993; Vollmann et al. 1996); and 4) moving average (Townley-Smith and Hurd 1973). Of particular interest are the nearest nei ghbor techniques, for which there are multiple variants. Nevertheless, the majority corresponds to modi fications of the Papadakis method (Atkinson 1969) which is based on the us e of one or more covariates obtained from averaging residuals of neighboring plots; these residuals are obtained from fitting a completely randomized or randomized complete block design. The original method uses the residual average of one pl ot to the right and one to the left; thus, known as EW adjustment. Several modifications are availabl e which consider different definitions of the covariates with 2, 4, 8 or even 24 adjacent plots (Brownie et al. 1993; Stroup et al. 1994; Vollman et al. 1996). Also, Bartlett (1978) s uggested performing an iterated Papadakis method that recalculate s residuals iteratively. A me thod similar to Papadakis is

PAGE 65

51 based on the original measurements and not on residuals is the M oving Average analysis. Here, a “local” mean is calculated from a number of adjacent plots and used as a covariate (Townley-Smith and Hurd, 1973). The second component of spatia l variability (local trends ), frequently identified as multiple patches on the field surface, can be modeled through the specification of an error structure that takes into a ccount some form of spatial correlation produced by the physical proximity among experimental units. Most of the technique s for modeling global trends (as presented above) assume that residuals or errors from the linear model are independent and identically distributed with a comm on variance. Under spatial correlation, the error variance-covariance matrix has nonzero off-diagonal elements, which are assumed to be a function of the distance among experimental units. Several correlated error models are available (Cressie 1993; Littell et al. 1996), allowing substantial flexibility in assumpti ons and covariance structures; nevertheless, different error structures tend to yield sim ilar results (Zimmerman and Harville 1991). The first-order separable autoregressive erro r structure has been us ed successfully in several agronomic and fo restry trials (Grondona et al. 1996; Gilmour et al. 1997; Qiao et al. 2000; Sarker et al. 2001; Costa e Silva et al. 2001; Dutkowski et al. 2002). This pattern considers two perpe ndicular correlations (one fo r row and the other for the column direction), and is equivalent to th e separable exponential ge ostatistical model. Sometimes a nugget parameter is included in th e error structure to model the expected variability that occurs if repeated measurem ents were made exactly at the same location (Cressie, 1993) allowi ng modeling of potential micros ite variability or measurement error.

PAGE 66

52 Finally, it is possible to fit global an d local trends simultaneously for which multiple combinations of techniques exist. Because many of these options yield similar results, it is more difficult to select an ade quate model, and the use of some statistical and graphical tools ( e.g. sample variograms) seems pa rticularly useful (Gilmour et al. 1997). Field testing of varieties is critical for the progress of genetic improvement programs, and is an expensive and time c onsuming activity. Theref ore, allocation of resources and the use of the best availabl e statistical techniqu es to obtain precise estimation of genetic parameters, such as heritability and breeding values, are critical. Due to the large number of varieties or treatments to be studied relatively large test areas are required. These areas usually have high am ounts of environmental heterogeneity that inflate the residual error varian ce. Hence, controlling for this variability is important, and requires the use of more sophisticated expe rimental designs, or the implementation of spatial statistical analyses. The goal of this study is to quantify effi ciencies of implementing a range of “ a priori ” and “ a posteriori ” statistical techniques to account for spatial variab ility, and also to discriminate among these techniques for parsimony and global performance for a broad set of conditions. In part icular, the selected array of techniques is compared to improvement in the precision for comparing and predicting treatment effects. This is achieved through simulations of field trials with sets of unrelated genotypes (or treatments) tested on a single site with environments having different patterns of heterogeneity (only patchy, only gradient, and both types). These patterns were generated using a polynomial function for the gradients and an AR1AR1 error structure for patches. The selected techniques included 1) modeling of global trends through the use of

PAGE 67

53 traditional experimental designs (complete randomized, randomized complete block, incomplete block and row-column) or polynomial models; 2) fitting AR1AR1 error structure (with and without nugge t) individually or in comb ination with some of the above specifications of globa l trends; and 3) implementati on of several variants of Papadakis and Moving Average methods. Materials and Methods Field Design and Simulation The simulations were based on a single trial of 2,048 plots arranged on a contiguous rectangular site of 64 rows a nd 32 columns with square spacing and no missing observations. For each site a total of 256 treatments (genotypes or varieties) with 8 replicates each were tested in plots composed by single plants. The linear model used to generate the simulated data for the response variable was yijk = + gk + Es(ijk) + Ems(ijk) 4-1 where yijk is the response of the plant located in the ith row and jth column of the kth treatment, is a fixed population mean which was set equal to 10 units, gk is the random treatment effect, Es(ijk) is the surface error (or st ructured residual) and Ems(ijk) corresponds to the microsite random error (or unstruc tured residual sometimes called nugget). The total variance for the above linear model is 2 T = 2 g + 2 s + 2 ms. For simplicity, but without loss of generality, 2 T was fixed to 1, and the variances for all surfaces were set to 2 g = 0.25 and 2 s + 2 ms = 0.75. The base experimental design simulated corresponded to an incomplete block design with 32 blocks in each of the 8 resolvable replications in an -lattice design (John and Williams 1995, p. 68), which had previously shown to be superior to completely

PAGE 68

54 randomized and randomized complete block de signs (Chapter 2). This experimental design was implemented in three surface patt erns: only patches (PATCH), only gradients (GRAD), and both types (ALL). These simula ted error surfaces included two main components: gradients and patches. The following third degree polynomial function using the x and y coordinates of each plot was used to generate the gradients: ) ( ) (2 2 ci ci ci ci ci ci iy x y x y x t 4-2 where xci and yci correspond to the centered position values for the ith plot ( i.e. xci = xi –x) located in column x and row y. Patches were generated using a first-order separable autoregressive error structure (AR1AR1) based on two perpendicular spatial correlations ( x and y) with an error variance-c ovariance matrix defined as Var (ei) = 2 s + 2 ms for diagonal elements 4-3 Cov (ei, ei’) = 2 s x hx y hy for off-diagonal elements 4-4 where hx = | xi xi’ |, hy = | yi yi’ |, i.e. the absolute distance in the row and column position, respectively, between two plots. The simulation consisted of the generation of independent surfaces over which different experimental designs were s uperimposed. Each surface was produced by selecting 5 parameters at random from uniform distributions (, x, y and 2 ms). The parameters were restricted to th e following ranges: 0 to 0.05 for and 0 to 0.0005 for 0.01 to 0.99 for the correlation parameters x and y, and 0.15 to 0.60 for 2 ms. Additionally, the two correlations were restrict ed so their absolute difference was smaller than 0.85 and 2 s was calculated from 2 s = 0.75 – 2 ms. Several layouts of the incomplete block design were generated using the software CycDesigN and randomly selected (Whitaker et al 2002). For each surface pattern 1,000 simulations were

PAGE 69

55 generated and stored for further analysis. Mo re details of the simulation process can be found in Chapter 2. Statistical Analysis For the prediction of treatment effects and es timation of other parameters of interest four sets of statistical analyses were cons idered based on modeling: 1) only global trends; 2) only local trends; 3) both components simultaneously; and 4) nearest neighbor methods. All linear models were fitted using the software ASREML (Gilmour et al. 2002), which estimates fixed effects, vari ance-covariance parame ters and predicts random effects by REML (Patterson and Thompson 1971). For those analyses that considered only gl obal trends it was assumed that the errors were independent and identica lly distributed; therefore, no spatial error structure was considered. To model the global trends or mean structure, cl assical experimental designs and explicit trend modeling were implemen ted. The classical designs or models corresponded to a completely randomized ( CR), randomized complete block (RCB), the original incomplete block de sign with 32 blocks (IB) which was employed to simulate the data, and a superimposed row-column (R-C ) design that consiste d of short rows and columns within each replicate. All effects with the excepti on of the overall mean were considered random (Table 4-1). The original -lattice design did not consider row and column effects; nevertheless, it is of interest to include this type of analysis because the short rows and columns allow for so me form of spatial modeling (Lin et al. 1993). Explicit trend modeling considered a set of continuous variables to describe polynomial regressions. The independent variables used in the study corresponded to the x and y coordinates of the individual plots, which were assumed fi xed. The models used are

PAGE 70

56 specified on Table 4-2. It is important to not e that the model Full-Poly was the one used to generate the gradients in the erro r surface simulations (see Equation 4-2). For modeling the local trends two variants of the AR1AR1 error structure were fitted with and without nugget effect. The va riant with nugget is sp ecified by Equations 4-3 and 4-4; and for the case of no nugget, the only difference is that the variance component 2 ms was assumed to be zero. These two variants were fitted alone or in combination with different form s of global trends: 1) the cl assical models specified in Table 4-1; and 2) the polynomial regressions from Table 4-2. Table 4-1. Linear models fitted for simula ted datasets using classical experimental designs to model global trends: co mplete randomized (CR), randomized complete block (RCB), incomplete block design with 32 blocks (IB), and rowcolumn (R-C) design. All model effects other than the mean were considered random. Model a CR yij = + gi + ij RCB yij = + repi + gj + ij IB yijk = + repi + block(rep)ij + gk + ijk R-C yijkl = + repi + col(rep)ij + row(rep)ik + gl + ijkl a g, treatment effect; rep, resolvable replicates; block(rep), incomplete block nested within replicate; col(rep), column nested within rep licate (short column); row(rep), ro w nested within replicate (short row); and residual. Table 4-2. Linear models fitted for simulated datasets using polynomial functions to model global trends: linear (Linear), quadratic (Red-Poly), and quadratic model with some interactions (Full-Poly ). All variables are assumed fixed and the treatment effect is random. Model a Linear yijk = + 0 xi + 1 yj + gk + ijk Red-Poly yijk = + 0 xi + 1 yj + 2 xi 2 + 3 yj 2 + gk + ijk Full-Poly yijk = + 0 xi + 1 yj + 2 xi 2 yj + 3 yj 2 xi + gk + ijk a x, longitudinal position of the plant; y, latitudinal position; g, treatment; and residual.

PAGE 71

57 For the nearest neighbor methods, a number of different variants of Papadakis (PAP) and Moving Average (MA) methods were implemented. The design effects (replicate, block, etc.) were not considered when these methods were implemented. The residuals (or deviations) used to implem ent the PAP method and its variants were computed as k ij ijy y ˆ, where yij is the observation of th e plot located in the ith row and jth column and ky is the simple treatment average (f itting a CR design). The covariates used to correct for differential yields ( Xij) were calculated using the average of these residuals from a number of neighboring pl ots. In the case of the MA method, the covariate was calculated using the original observations ( yij) instead of their residuals (ij ˆ). The following linear model, which considered the covariates as fixed effects and the treatment effects (gk) as random, was used yijk = + 0 X1,ij + 1 X2,ij + … + m Xp,ij + gk + ijk 4-5 where X1,ij, X2,ij, …, Xp,ij are p different covariates. All defin itions of covariates and linear models considered for this study are detailed in Figure 41. All Papadakis or Moving Average models were fitted with an error st ructure assuming independent errors. Also, an iterated PAP method was implemented for the models PAP-6, PAP-8 and PAP-11. Three iterations were implemented on these models using the residuals obtained from the previous iteration. In summary, a total of 35 different statis tical models where fitted for 3 difference surface patterns (PATCH, GRAD and ALL) each one with 1,000 simulated datasets totaling 105,000 sets of estimated parameters; additionally 24,000 sets were available from the iterative PAP models.

PAGE 72

58 The output from each of the fitted mode ls, datasets and surface patterns was compiled and summarized to obtain averages for variance component parameters. Also, empirical correlations between true and predicted treatment effect values (CORR) were calculated to compare different techniques and models. 1 1 2 1 1 2 2 1 1 2 1 2 2 1 2 1 4 3 3 4 1 2 1 1 1 1 1 1 1 1 1 1 1 1 2 1 2 1 1 2 1 2 3 1 4 2 2 4 1 3 2 5 1 6 4 3 3 4 6 1 5 2 1 1 PAP-1PAP-2PAP-3 PAP-4PAP-5PAP-6 PAP-7 / MA-1PAP-8 / MA-2PAP-9 / MA-3 PAP-10PAP-11 1 1 1 1 1 1 1 1 1 1 2 1 1 2 2 1 1 2 2 2 2 1 1 1 1 1 1 2 2 2 2 1 1 2 2 1 1 2 2 2 2 1 1 1 1 1 1 2 2 2 1 2 2 1 1 2 2 1 1 1 1 2 2 2 2 2 2 1 1 1 2 1 4 3 3 4 1 2 2 1 4 3 3 4 1 2 2 2 2 1 1 1 4 4 4 3 3 3 3 3 3 4 4 4 1 1 1 2 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 2 1 1 2 1 2 2 1 2 1 1 2 1 2 2 2 2 1 1 1 2 2 2 1 1 1 1 1 1 2 2 2 1 1 1 2 2 2 3 1 4 2 2 4 1 3 3 1 4 2 2 4 1 3 3 3 3 1 1 1 4 4 4 2 2 2 2 2 2 4 4 4 1 1 1 3 3 3 2 5 1 6 4 3 3 4 6 1 5 2 2 5 1 6 4 3 3 4 6 1 5 2 2 2 2 5 5 5 1 1 1 6 6 6 4 4 4 3 3 3 3 3 3 4 4 4 6 6 6 1 1 1 5 5 5 2 2 2 1 1 1 1 1 1 1 1 1 1 PAP-1PAP-2PAP-3 PAP-4PAP-5PAP-6 PAP-7 / MA-1PAP-8 / MA-2PAP-9 / MA-3 PAP-10PAP-11 Figure 4-1. Neighbor plots and definitions of covariates used in Papadakis (PAP) and Moving Average (MA) methods. Plots with the same numbers indicate a common covariate.

PAGE 73

59 Results Modeling Global and Local Trends Comparisons of the average correlati on values between true and predicted treatment effects showed cons iderable differences among linear models, error structures, and surface patterns with values that ranged from 0.85 to 0.90 (Figure 4-2). While this range is relatively narrow, small increments in values closer to 1 correspond to large improvements in terms of precision. In relati on to those models that assumed independent errors (ID), all designs yielde d higher average CORR values than CR. In general, the best model was R-C, but Full-Poly was better for surfaces with only global gradients (GRAD). In general, the models with explic it trends (Linear, Red-Poly and Full-Poly) were very good for GRAD surfaces and unsuccessful for sites with only local patches (PATCH). Nevertheless, for patchy surfaces pol ynomial models were slightly better than CR, which is an indication that a small porti on of the patches can be explained as trends. Approximately 98% and 95% of the fitted m odels converged for the autoregressive without nugget (AR1AR1) and with nugget (AR1AR1+) error structures, respectively. In GRAD surfaces, the results for the latter struct ure are not reported because very few simulations converged. Fo r all surface patterns, most of the failed convergences occurred in surfaces without patches, i.e. when the spatial correlations (x and y) were close to zero. This situation make s the covariance formula in Equation 4-4 equal to zero; hence, the vari ance of the error composed by 2 s and 2 ms as indicated in Equation 4-3 cannot be adequately partitioned. Once the errors were assumed correlated ( i.e. when the error structure was incorporated in the linear models) severa l changes occurred (Figure 4-2). First,

PAGE 74

60 AR1AR1 had for all models on surfaces ALL and PATCH larger average CORR values than model with independent error structur e. The largest improvement in CORR was found on PATCH surfaces, which was expected because in this case now the error structure is explicitly modeled. Also, the CORR values in GRAD surfaces almost did not change between AR1AR1 and ID error structures, w ith the exception of CR designs that had better CORR values than with ID. Once the nugget effect was incorporated, as with AR1AR1+, even larger CORR values were found. Comparisons of this erro r structure with the ot her two showed that differences between experiment al designs were even smaller. The maximum values for CORR were 0.90 and 0.88 with the Full-Poly model for ALL and PATCH, respectively. But it is important to note that this last model corresponds exactly to the one used to simulate the data; therefore, in this study it was used as re ference or base model for the maximum performance that can be achieved. As with CORR, the average estimated va riance components differed between error structures and models (see Ta bles 4-3, 4-4 and 4-5). And, as expected, within the same error structure, there was an inverse relations hip between the variance of the error and the average CORR value. For all designs, surfaces and error structures, the mean treatment variance component (2 g) after 1,000 simulations was close to its parametric value of 0.25; and, hence, unbiased. In addition, the estimate of the error variance (2 s or 2 s + 2 ms) was near its parametric value f 0.75 for a ll surface patterns and error structures (for models that partitions the erro r structure into replicates, bl ocks, rows and columns, these components are included in the total sum of 0.75).

PAGE 75

61 Figure 4-2. Average correlations between true and predicted treatment effects (CORR) in 3 different surface patterns for classi cal experimental design analyses and polynomial models fitted for the following error structures: independent errors (ID); autoregressive without nugget (AR1AR1); and autoregressive with nugget (AR1AR1+). Surface GRAD is not shown in the last graph because few simulations converged. The average variance components for the reference model from this simulation study (Full-Poly with AR1AR1+) for PATCH surfaces were close to their parametric constants, with the exception of x which was slightly overes timated (0.52 instead of 0.50). A similar situation was found in ALL, but in this case the error variance only sums up to 0.56 (instead of the parametric 0.75) b ecause part of the variability was explained by the polynomial fixed effects. Hence, fo r this surface explicit trends explained approximately 19% of the total variability. Within AR1AR1+ important differences were found when a model different than Full-Poly was used. In ALL surfaces, larger average x and y were obtained with values as high as 0.88. Because of the lack of trends in PATCH the spatial correlation for AR1 AR1+ CR R C B I B 32 R-C Li n e a r Red-Poly Fu l l -P o ly ID CR R C B IB 3 2 R-C Linear R ed -P o l y Full-Poly CORR 0.85 0.86 0.87 0.88 0.89 0.90 AR1 AR1 C R RCB IB 3 2 R -C L i n e ar Red Poly Ful l -Pol y ALL GRAD PATCH

PAGE 76

62 other models different than Full-Poly were close to 0.50. In the case of GRAD, the correlations for AR1AR1 should be close to zero, but in almost all models, slightly positive values were found with the exception of Full-Poly. This bias was particularly large for CR, possible a consequence of fa iling to model the underlying surface trends properly. Table 4-3. Average variance components fo r the surface pattern GRAD obtained from fitting the models in Tables 4-1 and 4-2 for the following error structures: a) independent errors (ID); and b) autoregressive without nugget (AR1AR1). All values are the means from 1,000 simulations and the true values were 2 g = 0.25 and 2 s + 2 ms = 0.75 for CR design. ID Model a g b rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.75 RCB 0.25 0.16 0.61 IB 32 0.25 0.16 0.05 0.56 R-C 0.25 0.16 0.04 0.02 0.55 Linear 0.25 0.60 Red-Poly 0.25 0.60 Full-Poly 0.25 0.56 AR1 AR1 Model g rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.19 0.20 0.74 RCB 0.25 0.16 0.07 0.07 0.61 IB 32 0.25 0.16 0.04 0.03 0.02 0.57 R-C 0.25 0.16 0.04 0.02 0.02 0.01 0.56 Linear 0.25 0.07 0.07 0.60 Red-Poly 0.25 0.06 0.07 0.60 Full-Poly 0.25 0.00 0.00 0.56 a CR, complete randomized; RCB, randomized complete block; IB 32 incomplete block with 32 blocks per replicate; R-C, row-column; Lin ear, linear regression; Red-Poly, re duced polynomial regression; FullPoly, full polynomial regression. b g, treatment; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row( rep), row nested within replicate; x, correlation in the x-direction; y, correlation in the y-direction; ms, microsite error; s, surface error.

PAGE 77

63 Table 4-4. Average variance components for the surface pattern ALL obtained from fitting the models in Tables 4-1 and 4-2 for the following error structures: a) independent errors (ID); b) au toregressive w ithout nugget (AR1AR1); and c) autoregressive with nugget (AR1AR1+). All values are the means from 1,000 simulations and the true values were 2 g = 0.25 and 2 s + 2 ms = 0.75 for CR design. ID Model a g b rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.75 RCB 0.25 0.17 0.60 IB 32 0.25 0.17 0.09 0.51 R-C 0.25 0.17 0.06 0.05 0.50 Linear 0.25 0.61 Red-Poly 0.25 0.60 Full-Poly 0.25 0.55 AR1 AR1 Model g rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.29 0.28 0.72 RCB 0.25 0.16 0.19 0.18 0.60 IB 32 0.25 0.16 0.02 0.16 0.13 0.58 R-C 0.25 0.16 0.04 0.03 0.12 0.12 0.53 Linear 0.25 0.20 0.19 0.60 Red-Poly 0.25 0.19 0.18 0.60 Full-Poly 0.25 0.15 0.14 0.55 AR1 AR1+ Model g rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.87 0.88 0.44 0.33 RCB 0.25 0.07 0.76 0.78 0.35 0.35 IB 32 0.25 0.06 0.00 0.79 0.79 0.42 0.29 R-C 0.25 0.07 0.00 0.00 0.75 0.77 0.35 0.33 Linear 0.25 0.73 0.72 0.41 0.22 Red-Poly 0.25 0.73 0.72 0.40 0.22 Full-Poly 0.25 0.54 0.51 0.36 0.20 a CR, complete randomized; RCB, randomized complete block; IB 32 incomplete block with 32 blocks per replicate; R-C, row-column; Lin ear, linear regression; Red-Poly, re duced polynomial regression; FullPoly, full polynomial regression. b g, treatment; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row( rep), row nested within replicate; x, correlation in the x-direction; y, correlation in the y-direction; ms, microsite error; s, surface error.

PAGE 78

64 Table 4-5. Average variance components fo r the surface pattern PATCH obtained from fitting the models in Tables 4-1 and 4-2 for the following error structures: a) independent errors (ID); b) au toregressive w ithout nugget (AR1AR1); and c) autoregressive with nugget (AR1AR1+). All values are the means from 1,000 simulations and the true values were 2 g = 0.25 and 2 s + 2 ms = 0.75 for CR design. ID Model a g b rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.75 RCB 0.25 0.02 0.73 IB 32 0.25 0.02 0.10 0.63 R-C 0.25 0.02 0.06 0.06 0.62 Linear 0.25 0.74 Red-Poly 0.25 0.74 Full-Poly 0.25 0.73 AR1 AR1 Model g rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.22 0.21 0.74 RCB 0.25 0.02 0.21 0.20 0.73 IB 32 0.25 0.01 0.02 0.20 0.18 0.71 R-C 0.25 0.02 0.04 0.03 0.17 0.16 0.66 Linear 0.25 0.22 0.21 0.73 Red-Poly 0.25 0.22 0.21 0.73 Full-Poly 0.25 0.21 0.20 0.73 AR1 AR1+ Model g rep block(rep) col(rep) row(rep) x y 2 ms 2 s CR 0.25 0.52 0.50 0.37 0.39 RCB 0.25 0.00 0.53 0.51 0.33 0.43 IB 32 0.25 0.00 0.00 0.51 0.49 0.36 0.39 R-C 0.25 0.00 0.00 0.00 0.51 0.50 0.33 0.42 Linear 0.25 0.52 0.50 0.37 0.40 Red-Poly 0.25 0.52 0.50 0.36 0.40 Full-Poly 0.25 0.52 0.50 0.37 0.40 a CR, complete randomized; RCB, randomized complete block; IB 32 incomplete block with 32 blocks per replicate; R-C, row-column; Lin ear, linear regression; Red-Poly, re duced polynomial regression; FullPoly, full polynomial regression. b g, treatment; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row( rep), row nested within replicate; x, correlation in the x-direction; y, correlation in the y-direction; ms, microsite error; s, surface error.

PAGE 79

65 Considering only models with independe nt errors, the partition of variance components showed larger values of the replicate variance co mponent in ALL and GRAD in comparison with PATCH. Also, the incomplete block variance components (from IB and R-C) were always smaller in GRAD surfaces. Because few simulations converged fo r GRAD surfaces, to draw inferences comparisons between AR1AR1 and AR1AR1+ can only be performed for surfaces ALL and PATCH. Failing to in corporate the nugget effect in these surface patterns produced an under estimation of the spatial co rrelation parameters for any experimental design. Another consequence was an incr ease in the magnitude of the variance components for other effects. For example, in ALL surfaces the replicate variance increased from 0.07 to 0.16, indicating that part of the variability that the error structured failed to capture was absorbed by the replicate effect. Nearest Neighbor Models Due to differences in the number of ne ighbor plots accounted for and covariates considered, a wide variety of average CORR values were f ound for the nearest neighbor methods (Figure 4-3), but in all cases, av erage CORR values greater than 0.86 were found. In general, models that considered mo re plots and/or covari ates produced larger CORR values. When compared with analyses with RCB designs, both the PAP and MA methods gave larger CORR values for PATCH and ALL, but for GRAD surface RCB was in general better than any nearest neighbo r methods. This situati on is not surprising because PAP and MA were originally desi gned to control for co rrelated errors (or patches) and not potential surface trends or gr adients. Nevertheless, as occurred before with the incorporation of autoregressive error structures, controlling for patches helped to

PAGE 80

66 model a small portion of the trends (see GRAD in Figures 4-2 and 4-3). Additionally, for all nearest neighbor methods studied, unbiased estimates of variance components were always obtained, as indicated by values clos e to 0.25 for all cases. Also, a proportional inverse relation was noted be tween the average error variance and the CORR values. Generally, PAP-6 and PAP-11 were superior to any of the MA me thods across all surface patterns (Figure 4-3), where PAP-11 was slight ly better, particularly for GRAD surfaces. Interestingly, the PAP-8 method that only c onsiders one covariate for the 8 surrounding plots was particularly good for GRAD. PAP-6, PAP-8 and PAP-11 were also fitte d iteratively using their previously estimated residuals to calculate new covariat es. For any given surface, a decrease in the average CORR and the treatment and error variance was noted in PAP-6 and PAP-11, this decrease was more abrupt in the last iteration (Table 4-6) and affected mostly surfaces with patches (ALL and PATCH). PAP-8 was almost not affected by the iterations showing stable statistics and unbiased variance components. A group of selected models was plotted in Figure 4-4. The first 4 models assumed that the errors were independent (ID) and the last 2 models corresponded to the best models for the autoregressive error structur es. ID models did not achieve the maximum potential CORR values that could be obt ained by using the Full-Poly with AR1AR1+. As indicated before, the cl assical design R-C was very good in ALL and GRAD, but less precise for PATCH surfaces than other m odels. PAP-6 and PAP-11 performed very well but had slightly lower CORR values than the maximum potential CORR obtained with the AR1AR1+.

PAGE 81

67 P A P 1 P A P 2 P A P -3 P A P 4 P A P -5 P A P -6 P A P -7 P A P -8 P A P -9 P A P 1 0 P A P 1 1 M A 1 M A 2 M A 3 CORR 0.85 0.86 0.87 0.88 0.89 0.90 ALL GRAD PATCH Figure 4-3. Average correlations between true and predicted treatment effects (CORR) in 3 different surface patterns for near est neighbor analyses fitted assuming independent errors. PAP: Papa dakis, MA: Moving Average. Figure 4-4. Average correlations between true and predicted treatment effects (CORR) in 3 different surface patterns for select ed methods: randomized complete block (RCB), row-column (R-C), and Papadakis (PAP) using the following error structures: independent errors (ID ), and autoregressive with nugget (AR1AR1+). RCBR-CPAP-6PAP-11Full-Poly Full-Poly CORR 0.85 0.86 0.87 0.88 0.89 0.90 ALL GRAD PATCH (ID) (ID) (ID) (ID) (AR1 AR1) (AR1 AR1+ )

PAGE 82

68 Table 4-6. Average correlation between true and predicted treatment effects (CORR) and average treatment variance component fo r iterated Papadakis methods PAP-6, PAP-8 and PAP-11 for the surface patterns ALL, PATCH and GRAD. ALL a PAP-6 PAP-8 PAP-11 Iteration CORR 2 g b 2 s CORR 2 g 2 s CORR 2 g 2 s 0 0.89 0.25 0.52 0.89 0.25 0.54 0.89 0.25 0.51 1 0.89 0.25 0.52 0.89 0.25 0.53 0.89 0.26 0.50 2 0.89 0.25 0.47 0.89 0.25 0.53 0.89 0.25 0.44 3 0.87 0.24 0.45 0.89 0.25 0.53 0.87 0.24 0.39 PATCH PAP-6 PAP-8 PAP-11 Iteration CORR 2 g 2 s CORR 2 g 2 s CORR 2 g 2 s 0 0.88 0.25 0.59 0.87 0.25 0.65 0.88 0.25 0.59 1 0.86 0.26 0.62 0.87 0.25 0.64 0.87 0.26 0.57 2 0.87 0.25 0.49 0.87 0.25 0.64 0.87 0.25 0.47 3 0.83 0.23 0.49 0.87 0.25 0.64 0.85 0.24 0.44 GRAD PAP-6 PAP-8 PAP-11 Iteration CORR 2 g 2 s CORR 2 g 2 s CORR 2 g 2 s 0 0.88 0.25 0.61 0.88 0.25 0.60 0.88 0.25 0.60 1 0.88 0.25 0.60 0.88 0.25 0.60 0.88 0.25 0.59 2 0.87 0.25 0.59 0.88 0.25 0.60 0.87 0.25 0.56 3 0.87 0.24 0.56 0.88 0.25 0.60 0.86 0.25 0.45 a ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. b g, treatment effect; s, surface error. Discussion Sites for field trials in agronomy or fore stry are particularly variable showing spatial heterogeneity that produces positive co rrelations between plots in close proximity. In this context, classical e xperimental designs that assume independent errors are still valid due to the randomization of treatments but they are not optimal. Some of the consequences are (Ball et al. 1993) 1) inflation of the mean square error (MSE); 2) MSE

PAGE 83

69 is not the minimum-variance estimator of the error; 3) overestimated standard error of treatment pairwise comparisons; 4) invalid hypothesis testing and confidence intervals; and 5) bias in other stat istics calculated from variance components estimates ( e.g. heritability and genetic correlations). Nevert heless, controlling for spatial variation by modeling the error stru cture or using neares t neighbor analysis helps to minimize these problems and to obtain more efficient analyses. A concern in this study was CORR values close to the maximum of 1, which was probably a consequence of using a large number of replicates in the field design. To study this, the first 500 simulated dataset from the ALL surface were fitted using only 2 contiguous replicates and the models from Tables 4-1 and 4-2 w ith all three error structures described previous ly. The results showed an av erage CORR for all surfaces that ranged from 0.67 to 0.71 (for ALL see Table A-8), which were lower than the ones obtained with 8 replicates (see Figure 4-2), but both numbers of replicates results were similar in range and behavior. Some relevant differences were a greater difficulty in achieving convergence and an increase d variability in the estimates. Modeling Global and Local Trends R-C was the best of all e xperimental design models f itted assuming independent errors (ID). This design which incorporates s hort rows and columns within replicates was particularly good for patches and in a less exte nt for gradients. These findings agree with other studies where R-C designs showed greater efficiency than RCB (Williams et al. 1999) and IB designs (Qiao et al. 2000). Hence, for all surface patterns tested in this study designing an experiment as R-C is recommended. Incorporation of the autore gressive error structure yi elded improved precision of treatment estimates as reflected in larger CORR values. The best model, as expected,

PAGE 84

70 corresponded to Full-Poly with AR1AR1+ which was the reference model used to simulate the surfaces. In contrast, when th e autoregressive error structure was fitted without the nugget effect only small decreases in average CORR values was noted, but considerable bias in some of the variance co mponents, particularly an underestimation of the spatial correlation parameters was found. For forestry breeding trials Dutkowski et al. (2002) also reported an overestimation when the nugget effect was not included in the model but in their case for the additive genetic variance component. Zimmerman and Harville (1991) indicated that in many cases this extra parameter might be unnecessary for most field experiments and could produce problems because of increased computation. But in this and several other studies, microsite error has been found to be necessary and often large relative to the total su rface error (Ball et al. 1993; Gilmour et al. 1997; Cullis et al. 1998; Costa e Silva et al. 2001; Dutkowski et al. 2002). All these results suggest that the incl usion of the nugget or microsite error might be important. Usually, portions of the gradient variability can be explained by the error structure (Zimmerman and Harville 1991). In this st udy, this situation was observed when the gradient was modeled incorrectly ( e.g. using Red-Poly instead of Full-Poly) under an autoregressive error structure fro GRAD surfaces some improvements on average CORR values over CR design were found. There is debate about the use of the design effects together with an error structure in the same model. These effects induce artificial discontinuities for adjacent experimental units on the boundaries (replicates or incomplete bloc ks). In the present study, once the error structur e was modeled through autore gressive with or without nugget, differences between experimental designs were considerably reduced,

PAGE 85

71 particularly for PATCH surfaces (see Figure 4-2). Nevertheless, in analyses of real datasets, some authors using au toregressive errors together with design effects have been successful in increasing precision of tr eatment effect predictions (Sarker et al. 2001; Dutkowski et al. 2002). The concept behind modeling global trends is to represent the field gradient that is generated by gradual change s in topography and soil char acteristics. Polynomial functions were used in this study, which unde r independent errors proved to be suitable only on those simulated surfaces with dominant gradients (ALL and GRAD). Importantly, once the correct autoregressive error structure was incorporated (AR1AR1+) no differences were found between models with different polynomial functions or design effects. Differences were only noted in the GRAD surface were FullPoly was clearly better than a ny other alternative. In severa l studies, the in corporation of fixed or random effects to model global tre nds together with the autoregressive error structure has resulted in im proved precision of treatment estimates or predictions (Brownie et al. 1993; Brownie and Gumpertz 1997; Gilmour et al. 1997). Here, trend or gradient was modeled using polynomial functi ons, but other options such as smoothing splines (Verbyla et al. 1999) or differencing the data (Cullis and Gleeson 1991) have proved useful. In this study, each simulated dataset was f itted in an automated fashion. Due to the variability between trials, a general model cannot adequately fit all experiments. For routine analyses it is recomme nded, as suggested by Gilmour et al. (1997), that for each individual datasets several variants of the au toregressive spatial model must be tried, and the linear model extended by the inclusion of other design factors, polynomial functions,

PAGE 86

72 or other fixed or random effects to model th e gradient. Also, mode l selection should be always accompanied by the use of diagnostic tools, such as variograms (Cressie 1993) or the use some of the available mixed model information criteria. Nearest Neighbor Models The findings in this study indicated th at considerable improvements can be achieved with PAP or MA methods in controllin g for spatial variation. As expected, the best results were obtained for patchy surf aces. For the same covariate definition, PAP methods were always better than MA. This situation demonstrates that, as done with PAP, it is necessary to eliminate the treatme nt effect from the response variables. The performance of PAP methods agrees with ot her studies in which they were always superior to RCB designs, gene rally better than IB desi gns (Kempton and Howes 1981; Brownie et al. 1993; Vollmann et al. 1996), and only slightly in ferior to models that fitted the autoregressive error structur e (Zimmerman and Harville 1991; Brownie et al. 1993; Brownie and Gumpertz 1997). The best nearest neighbor methods corre sponded to PAP-6 and PAP-11, which are based on 4 and 6 covariates respectively, and yi elded better results than the original PAP method based on a single covariate (PAP-1 and PAP-3). This agrees with other studies where the inclusion of more than one cova riate produced better results (Kempton and Howes 1981; Vollmann et al. 1996). With extra covariates it is possible to use more information about the neighbor plots; theref ore, improving efficien cy. Nevertheless, for different situations, the ideal set of covariates needs to be investigated. PAP methods and other near est neighbor methods not studied here have the advantages of simplicity and flexibility. Th ey are simple because only the assumption of independent errors is required so computa tional requirements are mi nimal, and they are

PAGE 87

73 flexible because several covariates with va rying number of plots can be defined and studied. Nevertheless, these methods are not fr ee of difficulties. Testing the inclusion of covariates in a model is difficult because it is not clear how ma ny degrees of freedom should be considered for each covariate due to the fact that a previous estimate of the treatment mean was used. As point of refe rence, Pearce (1998) in a simulation study recommended considering 2 degrees of freedom for each covariate. Also, it is not clear how to treat covariates obtained from border plots, or what to do if no border plots exist. Ideally, measurements outside the test area should be available as plots that received a treatment with an average response (Scharf and Alley 1993) Nevertheless, it is common to use only existing plots to cal culate covariates (Wilkinson et al. 1983), or replace nonexistent plots by their interior complement (Magnussen 1993). A similar situation occurs with missing information that generates an inte rnal border. Another re ported difficulty is that these methods appear to be more se nsitive to outliers (Kempton and Howes 1981) and to the presence of abrupt changes in the error surface (Binns 1987; Pearce 1998); and in many cases it is not clear what are the ideal solutions for these problems. The iterative PAP was first suggested by Bartlett (1978) to improve efficiency. Wilkinson et al. (1983) claimed that there is a considerable positive bias for testing significance of treatment effects for a single covariate iterative PAP method; nevertheless, Pearce (1998) in another study found no bias, and showed that some PAP methods were more effective after iteration. The results in this study for iterated PAP methods were not favorable producing a reductio n in the efficiency as more iterations were executed. Given the results, non-iterated PAP methods appear to be the best option. Nevertheless, is important to consider that for the present simulations a relatively large

PAGE 88

74 number of replicates per treatment were available ( i.e. 8) that allowed reasonable initial estimates of treatment mean s to calculate covariates. Apparently, the PAP methods can be im proved by including model effects to incorporate gradients or trends (which is where these methods failed to outperform other options, see Figure 4-4). Design effects, as replicates or blocks have been used in some studies together with some nearest ne ighbor covariates (K empton and Howes 1981; Scharf and Alley 1993) with varying outcomes. But, as indicated before, the use of design effects is not recommended because th ey create artificial discontinuities for neighboring plots that belong to different blocks. Polynom ials functions or smoothing splines do not have this problem because they produce a continuous gradient; hence, they are recommended for modeling gradients when employing nearest neighbor methods to account for patchiness. Usually, in spatial analysis only positive co rrelations due to physical proximity are considered. But, in many real situations a negative correlation can be generated by competition between plots belonging to differe nt treatments. This competition can disrupt the spatial correlation (Kempton and Howes 1981) and might require explicit modeling. However, this negative correlation is only pr esent once competition for resources (light, water and nutrients) is important, which might take several years to occur. Because the simulations in this study did not consider these effects so me caution must be taken in situations with high levels of them. The use of spatial techniques discussed he re does not invalidate classical design of experiments. As Sarker et al. (2001) suggest, it is important to design the layout of an experiment according to a preferred design ( e.g. IB), which should be first analyzed as

PAGE 89

75 designed, but further spatial models should be fitted to capture other relevant information (or noise) present in the field. For any design or analysis, tr eatment randomization is still critical to neutralize the eff ect of spatial correlations a nd provide unbiased estimates and to validate hypothesis testing (G rondona and Cressie 1991; Brownie et al. 1993). Also, specific design of experiments in the pres ence of spatial correlation might yield improvements, particularly in balancing trea tment pairwise comparisons. Little research in this topic has been reported, and for ex ample, Williams (1985) described an algorithm to improve IB designs by consid ering their spatial arrangement. The results from this study concentrated in single-site analys es. Combining several sites and other sources of information togeth er can yield to impr oved estimation of the genetic parameters. Usually, for multiple-site analyses each test is studied individually to determine its significant model effects and error structure, and later, all sites are combined in a unique model using their singl e-site specifications obtained previously. Specialized software is required to fit this complex mixed linear model, which commonly presents difficulties in getting variance compone nts that converge. In contrast, the use of nearest neighbor models on multiple-site analys es is considerable easier, particularly because the error structure is simple, and it only requires the definition of site specific covariate parameters. Nevertheless, furthe r studies in these topics are required. Conclusions The main findings of this study indicate d that an adequate selection of the experimental design can pr oduce considerable improveme nts in the prediction of treatment effects and better design options than the widely used RCB are available for field trials in agronomical and forestry studies. Particularly among models assuming independent errors, R-C designs gave the best results and IB designs of 32 incomplete

PAGE 90

76 blocks were almost as good. Nevertheless, in PATCH surf aces these designs did not perform as well as autoregressive error st ructures or nearest neighbor methods. Also, those designs that modeled global trends ( i.e. polynomial functions) performed well in surfaces with a dominant gradient (G RAD and ALL) but poorly in PATCH. The incorporation of a separable autoregr essive error structur e with or without nugget yielded the highest CORR values among all models and methods tried for all surface patterns. Here, the inclusion of de sign effects (replicates and blocks) or polynomial covariates were only useful wh en the nugget effect was not included. Nevertheless, in this last situation biased spatial correlations parameters were found. Promising results were obtained with th e use of nearest neighbor techniques, particularly for surfaces with patches, but in surfaces with gradients some minor improvements were also found. PAP methods were always better than MA, and they were almost as good as methods that modele d the error structure, particularly those variants that considered more plots and/or covariates. PAP11 was the best alternative followed by PAP-6, but the use of the latter is recommended because of its simplicity. The use of an iterated PAP method did not produce improvements in this study, generating deteriorated estimates. In summary, if simple analyses ar e preferred, R-C and IB designs with independent errors should be preferred. For fu rther improvements, it is recommended that some form of PAP methods be used, but seve ral covariates must be tried and tested. Finally, spatial analysis incor porating error structures are pr omising, and they should be used whenever possible, but the computati onal requirements and so me uncertainty about the correct procedures and testing may limit its practical use.

PAGE 91

77 CHAPTER 5 POST-HOC BLOCKING TO IMPROVE HERITABILITY AND BREEDING VALUE PREDICTION Introduction Genetic testing of provenances, varieties and families in forestry and agricultural crops is usually a long process and frequently one of the most expens ive activities of any genetic improvement program (Zobel and Talb ert 1984, p. 232). Appropr iate selection of an experimental design and st atistical analysis can produce considerable improvement in the prediction of breeding values and efficient use of available resources. Traditionally, in forest trials the randomized complete block (RCB) design has been favored. RCB design is effective when within replicate (or block) variability is relatively small, a situation that is rare in fo rest sites, or occurs only with relatively small replicates (Costa e Silva et al. 2001). Since a large number of genetic entities are usually tested together, a good alterna tive is to implement incomplete block designs (IB) which subdivide the full replicate in to smaller compartments ( i.e. incomplete blocks) that tend to be more homogeneous. For these designs several authors have reported greater efficiencies in comparison to RCB (Fu et al. 1998; Fu et al. 1999; see Chapter 2). It is also possible to simultaneously implement tw o-way blocking using row-column designs (R-C). These R-C designs are described in detail by John and Williams (1995, p. 87) and often yield even greater effici encies than IB designs (Lin et al. 1993; Qiao et al. 2000; see also Chapter 2).

PAGE 92

78 Fitting the best linear model that adequate ly describes the trial data together with the appropriate specification of random and fixed effects is critical. In the literature several techniques are availabl e for improving statistical anal yses, such as 1) spatial models (Gilmour et al. 1997; Chapter 4); 2) neares t neighbor methods (Vollmann et al. 1996; Chapter 4); and 3) inclusion of covariat es and other factors to deal with missing values (Cochran 1957) or nuisance effects (Gilmour et al. 1997). These are examples of “ a posteriori ” analyses. In these options, an exte nded linear model different than the original experimental design is fitted, which aims to model elements that were unnoticed or not controlled at test establishment. Another “ a posteriori” technique, called post-hoc bl ocking, can also be easily implemented. This method consists of supe rimposing a blocking structure on top of the original field design and a linear model is fitte d as if the blocking effects were present in the original design. Usually, an IB design is fitted over a simpler design ( e.g. RCB), but an R-C can also be superimposed. This te chnique was originally proposed by Patterson and Hunter (1983) as a tool to inexpensiv ely evaluate the efficiency of potential experimental designs. Nevertheless, seve ral authors have used post-hoc blocking successfully to increase heritabi lity and the precision of prediction of genetic effects for final analyses (Ericsson 1997; Dutkowski et al. 2002; Lopez et al. 2002). The primary problem in implementing post-hoc blocking is determining the characteristics of the incomplete blocks to be superimposed. If large incomplete blocks are used, the blocks will con tinue to be highly heterogeneous. On the other hand, small incomplete blocks should considerably re duce the residual vari ability by capturing smaller areas (or patches) of the environm ental surface. But if block size becomes too

PAGE 93

79 small, fewer treatments will occur together in the same block; therefore, the standard error of their difference can be considerab ly larger. Also, according to Ericsson (1997), with small blocks part of the treatment (o r genetic) variability can be absorbed by the block variance due to an increased chan ce that two good (or bad) genotypes occur together in the same block. Another problem is the shape of the incomplete block; nevertheless, general guidelines used for de sign of experiments can be followed. For example, a shape as close to a square as po ssible, and orienting th e blocks main axis perpendicularly to the principal field grad ient should be preferred (Zobel and Talbert 1984, p. 248). The present study used simulated single -site clonal trials and 3 different environmental patterns to understand conseque nces of and define fu ture strategies for using post-hoc blocking in es timating heritabilities and predicting breeding values. Specifically, the objectives were to 1) st udy the use of post-hoc blocking for several experimental designs, surface patterns and num bers of ramets; 2) verify a potential reduction in the genetic variance as the incomp lete blocks become smaller; 3) compare the performance of different strategies or cr iteria for selecting an appropriate blocking structure; and 4) disc uss statistical and practical issues related to implementing post-hoc blocking. Materials and Methods Several single site clonal trials were si mulated based on a rectangular grid of 64 rows and 32 columns with square spacing and no missing observations (see Chapter 2 for details). Briefly, 256 clones each with 8 ra mets were “planted” in single-tree plots according to several experimental plans. The simulated environmental or surface patterns were based on two main elements: a gradient or trend generated with a third degree

PAGE 94

80 continuous polynomial function, an d patches that were incor porated through a covariance structure based in a first-order sepa rable autoregressive process or AR1AR1 (Cressie 1993; Gilmour et al. 1997). The 3 surface patterns considered were 1) only patches (PATCH); 2) only gradient s (GRAD); and 3) both co mponents (ALL). The variance structure was set for a completely randomized design to have a si ngle-site individual broad-sense heritability ( HB 2) of 0.25 originating from a to tal variance fixed at 1.0 and a clonal variance of 2 clone = 0.25. The simulated designs were a randomized complete design with 8 resolvable replicates (RCB), in complete block designs with 4, 8, 16 and 32 incomplete blocks within each of the 8 full repl icates (IB 4, IB 8, IB 16 and IB 32), and a row-column design with 16 rows and 16 co lumns (R-C). For each design, several implementations were generated usin g the software CycDesigN (Whitaker et al. 2002) that produces -designs. For each combination of surface pattern and design type 1,000 different datasets were generated. The software ASREML (Gilmour et al. 2002) was used to fit all mixed linear models and to obtain restricted maxi mum likelihood (REML) variance component estimates and best linear unbiased predicti ons (BLUP) of clonal values (Patterson and Thompson 1971). To compare alternatives, seve ral statistics were calculated: summary statistics of estimated variance components, average empirical correlation between true and predicted clonal BLUP values (CORR) a nd individual single-s ite heritabilities (2ˆBH ). Finally, for each linear model and data set the log-likelihood value (logL) was recorded together with its average standard error of the differe nce (SED) between any two clones.

PAGE 95

81 Table 5-1. Linear models fitted for simula ted datasets using classical experimental designs to model global trends over all surface patterns: ra ndomized complete block (RCB), incomplete block design with x blocks (IB), and a row-column (R-C) design. All model eff ects other than the mean were considered random. Model a RCB yij = + repi + clonej + ij IB x yijk = + repi + block(rep)ij + clonek + ijk R-C yijkl = + repi + col(rep)ij + row(rep)ik + clonel + ijkl a clone, clone effect; rep, resolvable replication; block(rep), incomplete block nested within replication; col(rep), column nested within replication (short colu mn); row(rep), row nested within replication (short row); residual. Figure 5-1. Experimental layout for randomized comp lete block design simulations with 8 resolvable replicates (left) and part itioning of one replicate for incomplete block design layouts (right ). All simulations had 2 56 unrelated clones with 8 ramets per clone per site for a total of 2,048 trees.

PAGE 96

82 The first stage in this study consisted of fitting the original experimental designs (without post-hoc blocking) using the correspon ding linear models (see Table 5-1). Later, by employing only the RCB designs for all su rface patterns, post-hoc blocking was implemented by superimposing 4, 8, 16 and 32 incomplete blocks of sizes 8 x 8, 8 x 4, 4 x 4, and 4 trees x 2 trees, respectively, and a R-C design with short rows and columns of length 16 (see Figure 5-1). Th ese “new” post-hoc designs were fitted using the corresponding linear model from Table 5-1. Statistical compar isons of individual heritabilities between original designs and post-hoc blocking were made using a z-test. In a second stage, the first 500 datasets generated for RCB designs were used to perform additional post-hoc blocking studi es. The same post-hoc blocking designs previously used were fitted and two extra inco mplete block designs we re incorporated to examine the effects of very small incomple te blocks: a layout with 64 (IB 64) and 128 (IB 128) blocks were superimposed with bloc k sizes of 2 x 2 and 2 x 1 respectively (see Figure 5-1). The linear models from Table 51 were fitted to all datasets, and the previously described statistics were also calculated. Additiona lly, for each of the 500 datasets, a subset of 2 contiguous replicates was selected at random constituting a “new” trial based on only 2 ramets per clone. This re duced trial size was used because of interest in studying the potential eff ects of using post-hoc blocki ng in the estimation of the variance components and other st atistics for trials with fewer replications. For these datasets, the same incomplete block and rowcolumn designs described previously were fitted (IB 4, IB 8, IB 16, IB 32, IB 64, IB 128 and R-C), and summary statistics were calculated.

PAGE 97

83 In practical situations, seve ral blocking alternatives (var ying numbers of blocks and block shapes) are tested during an alysis, and a final model must be selected. In this study, the performance of log-likelihood and SED d ecision criteria was studied and compared with individual heritabilities. The first crite ria, logL, selects the blocking structure that maximizes the log-likelihood of the mixed linear model (Lopez et al. 2002). For SED, the model with the smallest average standard de viation of the difference between treatments is selected. The procedure consisted of select ing and recording the be st model fitted in the second stage using both of these criteria fo r each of the 500 data sets and experimental designs. Here, frequency tables were obtained with all desi gns together, and only for IB designs. Results When results from fitting the original experimental designs were compared to posthoc blocking designs of the same block size and orientation, small differences were found for all statistics. The largest difference for average 2ˆBH was 0.003 units (Table A-9). Further, z-tests comparing i ndividual heritabilities between the original design and posthoc blocking for each design type and su rface pattern indicated that no significant differences existed (experiment-wise = 0.05), with the single exception of R-C designs on PATCH surfaces. Also, when standard deviations for 2ˆBH were compared between original and post-hoc blocking analyses, the differences in magnitude were very small and probably due to sampling variation (Table A-9). The average clonal variance did not differ between original and posthoc blocking, and a negligible decrease in the precision of this component was found for post-hoc bloc king. A similar situation of very little

PAGE 98

84 difference between original experimental designs and post-ho c blocking was also observed for average correlati on between true and predicted clonal values (Figure 5-2). Average variance components obtained in the second stage empl oying even smaller post-hoc blocks for all designs and surface patter ns for the case of 8 ramets per clone are shown in Table 5-2. The clonal variance was alwa ys close to its parame tric value of 0.25, with no reduction in its average value as the size of the block decrea sed, as postulated by Ericsson (1997). Also, the bloc k variance increased and the e rror variance decreased as the number of incomplete blocks increased. C onsiderable differences were noted in block variance components be tween IB 4 and IB 128, indicat ing a large influence of the experimental design used on partitioning total variance. RCBIB 4IB 8IB 16IB 32R-C CORR 0.82 0.84 0.86 0.88 0.90 ALL Original ALL Post-hoc PATCH Original PATCH Post-hoc GRAD Original GRAD Post-hoc Figure 5-2. Average correlati on between true and predicte d clonal values (CORR) for original and post-hoc bloking analys es on ALL, PATCH and GRAD surface patterns for simulated surfaces with 8 ramets per clone.

PAGE 99

85 Table 5-2. Average variance components and individual broad-sense heritability for all surface patterns obtained from fitting the models in Table 5-1 from simulated datasets with a randomized complete block designs with 8 ramets per clone per site. ALL a Model b clone c rep block(rep) col(rep) row(rep) error 2ˆ B H RCB 0.25 0.17 0.60 0.296 IB 4 0.25 0.16 0.05 0.56 0.310 IB 8 0.25 0.17 0.07 0.54 0.319 IB 16 0.25 0.17 0.08 0.53 0.323 IB 32 0.25 0.17 0.09 0.51 0.329 IB 64 0.25 0.17 0.10 0.50 0.337 IB 128 0.25 0.17 0.10 0.50 0.336 R-C 0.25 0.17 0.06 0.05 0.50 0.336 PATCH Model clone rep block(rep) col(rep) row(rep) error 2ˆBH RCB 0.25 0.02 0.73 0.256 IB 4 0.25 0.01 0.03 0.71 0.263 IB 8 0.25 0.01 0.05 0.68 0.269 IB 16 0.25 0.02 0.07 0.67 0.274 IB 32 0.25 0.02 0.10 0.63 0.284 IB 64 0.25 0.02 0.13 0.60 0.295 IB 128 0.25 0.02 0.15 0.58 0.304 R-C 0.25 0.01 0.06 0.06 0.62 0.288 GRAD Model clone rep block(rep) col(rep) row(rep) error 2ˆBH RCB 0.25 0.17 0.61 0.292 IB 4 0.25 0.15 0.05 0.57 0.305 IB 8 0.25 0.16 0.05 0.56 0.308 IB 16 0.25 0.16 0.05 0.56 0.308 IB 32 0.25 0.16 0.05 0.56 0.308 IB 64 0.25 0.16 0.06 0.54 0.314 IB 128 0.25 0.17 0.06 0.55 0.313 R-C 0.25 0.16 0.04 0.02 0.55 0.312 a ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. b CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. c rep, resolvable replicate; block(rep), incomplete bloc k nested within replicate; col(rep), column nested within replicate; row(rep), row nest ed within replicate; clone, clone; residual.

PAGE 100

86 R CB IB 4 IB 8 IB 16 IB 3 2 IB 64 IB 1 2 8 R C CORR 0.85 0.86 0.87 0.88 0.89 0.90 RCB IB 4 IB 8 IB 16 IB 3 2 IB 64 IB 1 28 R C Average Heritability 0.24 0.26 0.28 0.30 0.32 0.34 ALL PATCH GRAD a b Figure 5-3. Average individual he ritability (a) and correlation between true and predicted clonal values (b) calculated over 500 simulations with a randomized complete block designs with 8 ramets per clone per site for each su rface pattern and post-hoc design. For average individual heritability (Figur e 5-3a) IB 128 was the best design across all 3 surface patterns, but IB 64 and R-C were relatively cl ose. In fact, considerable improvements in 2ˆBH can be achieved with post-hoc blocking using better designs; average 2ˆBH values of up to 0.34, compared with a base value of 0.25 for the completely randomized design were obtained for smaller bl ock sizes. Nevertheless, a better statistic to use is the average correlat ion between the true and predic ted values (CORR) because it evaluates how well a post-hoc blocking desi gn predicts clonal breeding values and, hence, genetic gain from clonal selection. The trends between 2ˆBH and CORR differ considerably (see Figure 5-3), and a decrea se in CORR was present for those designs

PAGE 101

87 with very small blocks. This decrease was particularly noteworthy for IB 128. Hence, according to average CORR, the best post-hoc design for ALL and PATCH was R-C, and for GRAD surfaces IB 8. Post-hoc blocking w ith IB 8, IB 16 and IB 32 were better than RCB for all surface patterns a nd very good for ALL and GRAD. Table 5-3. Relative frequenc y of best post-hoc blocking f its from 500 datasets for each post-hoc design and surface pattern with 8 and 2 ramets per clone per site according to the log-likelihood (logL) and average standard deviation of the difference (SED). The values in bold co rrespond to the most frequent designs. 8 ramets / clone / site logL SED Type a ALL b PATCH GRAD ALL PATCH GRAD RCB 0 % 0 % 0 % 0 % 0 % 0 % IB 4 2 % 0 % 49 % 0 % 0 % 24 % IB 8 38 % 4 % 51 % 23 % 3 % 73 % IB 16 11 % 4 % 0 % 9 % 1 % 0 % IB 32 4 % 19 % 0 % 9 % 18 % 0 % IB 64 1 % 17 % 0 % 4 % 21 % 0 % IB 128 0 % 11 % 0 % 0 % 18 % 0 % R-C 44 % 45 % 0 % 54 % 38 % 3 % 2 ramets / clone / site logL SED Type ALL PATCH GRAD ALL PATCH GRAD RCB 1 % 1 % 1 % 4 % 5 % 5 % IB 4 7 % 2 % 28 % 6 % 3 % 15 % IB 8 24 % 8 % 33 % 15 % 9 % 23 % IB 16 9 % 4 % 3 % 6 % 6 % 8 % IB 32 7 % 16 % 2 % 13 % 15 % 6 % IB 64 4 % 15 % 8 % 8 % 17 % 7 % IB 128 3 % 9 % 4 % 6 % 12 % 7 % R-C 45 % 45 % 21 % 43 % 33 % 28 % a RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. b ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches.

PAGE 102

88 The results obtained for 2 ramets per clone are not shown explicitly, but followed similar trends to the analyses with 8 ramets per clone. As expected a reduction in CORR was noted ( e.g. CORR ranged from 0.64 to 0.70) t ogether with an increase in their variability. However, no bias in the clonal variance estimates was found, and the best designs were same as with 8 ramets per clone. Relative frequencies of the best fits for th e different selection cr iteria are shown in Table 5-3 for each design type, surface patt ern and for 2 and 8 ramets per clone. Both logL and SED criteria were similar show ing an analogous trend to average CORR (Figure 5-3b). Interestingly, with 8 ramets per clone for both criteria, the RCB design was never selected as the best fit, which indi cates that post-hoc blocking always yielded improved precision for all statistics ( HB 2, CORR, logL and SED). For 2 ramets per clone, all designs were selected at least once, and differences between the most and least frequently selected design we re not as disparate as with 8 ramets per clone. This phenomenon resulted from higher variability in the parameters estimates product of a reduced number of observations (see Tables 3-4 and A-8). Discussion Differences between designed and post-hoc blocking were almost negligible in terms of average individual heritabilities, CORR values and variance components for all surface patterns and number of ramets; also no decrease in precision was noted. In other studies, Dutkowski et al. (2002) reported reductions of up to 4% in the error variance using the logL criteria to analyze gene tic forestry trials. In contrast, Qiao et al. (2000) applied post-hoc blocking to several agri cultural trials where they found only small improvements. It appears that post-hoc blocking is useful when there is environmental heterogeneity (in form of patches or gradient s) that can be contro lled through blocking.

PAGE 103

89 To demonstrate this, average CORR values we re calculated from subsets with high and low spatial correlations (Figure 5-4), where low correlations correspond to surfaces with patch of small size. As expected, differen ces between blocking structures were not important under low levels of spatial correlati on, but were for high levels. In a related study, detailed analysis of su rfaces indicated that smaller mi crosite variability and larger spatial correlations ( i.e. more homogeneous sites) show ed larger average heritability values and, therefore, better pred iction of clonal values (Chapter 3). R CB IB 4 I B 8 IB 16 IB 32 IB 6 4 IB 1 2 8 R -C CORR 0.82 0.84 0.86 0.88 0.90 0.92 PATCH High PATCH Low ALL High ALL Low Figure 5-4. Average correla tion between true and predic ted clonal values (CORR) calculated for ALL and PATCH surface patt erns and all post-hoc designs for sets with High ( x, > 0.7 and y > 0.7) and Low ( x < 0.3 and y < 0.3) spatial correlations. The datasets used corresponded to simulated surfaces with a randomized complete block design and 8 ramets per clone per site.

PAGE 104

90 Ericsson (1997) indicated that post-hoc anal ysis with small incomplete blocks can reduce the genetic variance by confounding the bl ock and genetic effects. This problem was not found in the present simulations for any of the surface patterns or level of number of ramets per clone. This difference was probably due to the fact that in this study incomplete block effects were treated as random in comparison with Ericsson’s (1997) who incorrectly assumed them as fixed. Assuming block effects random is extremely important in mixed linear models, because the need to specify a correlation between observations that be long to the same block. For optimal block size during experimental design, Williams et al. (2002, p. 131) recommended as a general rule using the s quare root of the numb er of treatment or entities, which for the present study is 256 = 16, i.e. the IB 16 design. This is an adequate option for post-hoc blocking too if the dominant surface patterns are unknown; however, when the surface has a dominant grad ient, larger post-hoc blocks are to be preferred ( e.g. IB 8), and for dominant patches sm aller post-hoc blocks should be used ( e.g. IB 32). On the other hand, the two-wa y blocking produced by R-C designs yielded some of the best statistics for all surfaces, except for GRAD where they were marginally inferior to IB 8; hence, for the simulated c onditions of the present st udy, it appears that in general the post-hoc R-C design should be pref erred over IB designs. The use of R-C in post-hoc blocking over an RCB design has not been reported previously in the literature. For the present study only one variant was an alyzed (with small row and columns within replicate), and further studies are required to understand the effects of other R-C layouts. The use of individual broad-sense heritability as a selection criterion for the best fit for choosing among post-hoc blocking designs can be deceptive (Figure 5-3a). The

PAGE 105

91 optimal criterion for genetic studies should be CORR, which is not available from many mixed model software packages. Neverthele ss, average logL and SED values were always proportional to CORR for all designs and surface patterns; hence, their use is recommended. However, we believe logL s hould be preferred, because SED appears to be more sensitive to imprecision in the es timation of clonal variance for individual dataset analyses. The use of post-hoc blocking is not free from controversy. Originally, this tool was conceived to investigate if a different bl ocking structure other than the originally implemented experimental design would be better suited for future experiments (Patterson and Hunter, 1 983). According to Pearce et al. (1988, p. 294) superimposing a different blocking system will bias the es timation of the error variance because the original randomization does not work properl y for alternative bloc king structures. In addition, one of the desirable properties of de signing an experiment as IB is to obtain exactly (or approximately) the same preci sion on pairwise treatment comparisons throughout the experiment, which is a situati on that post-hoc blocki ng does not ensure. Another important problem is the subjectivity of post-hoc blocking in selecting the size and shape of the blocks. Here, Pearce (1976) points that re-analyzing the “data until something appears that can be declared signi ficant” is of high ri sk, and can yield to biased estimates. On the other hand, arguments in favor of post-hoc blocking indicate that it is always difficult to foresee all the potential fact ors that can affect a particular design, and unplanned events occur during e xperimentation that require th e original statistical linear model to be modified in or der to control for new factor s (Federer 1988). Some of the

PAGE 106

92 frequently used post-hoc techniques include th e use of covariates to deal with nuisance effects and missing values and/or modeling of gradients in spatial an alysis. Ignoring this type of information corresponds to an ineffi cient use of relevant knowledge. In addition, the negative effects of post-hoc blocking in unbalanced pairwise treatment comparisons is usually overcome by the reduction of the erro r variance that occurs (for an example see Qiao et al. 2000). Besides, for genetic experiment s, a balanced pair wise comparison of genetic entities is not as relevant as an increased precision of the estimation of genetic parameters or prediction of breeding values, which for simulations studies is adequately summarized by the correlations between true and predicted genetic values. It is important to recall that optimal resu lts will always be obt ained with the proper planning and designing of experi ments before field establishment. For this, specific literature and software are available ( e.g. CycDesigN) to accommodate for the majority of experimental situations. Ne vertheless, as experiments ar e still established with poor local control ( i.e. CR or RCB), they can benefit from post-hoc blocking with practically no additional cost.

PAGE 107

93 CHAPTER 6 CONCLUSIONS Clonal testing is becoming a very important activity in breeding and deployment of many commercial forestry species. Because it is a relatively new practice, several aspects need to be understood and solved. Genetic testing of clones is one area for which guidelines are lacking for adequate characteri stics of clonal field trials. In addition, research done on testing of progenies using seed lings from half or full-sib families is only partially applicable to clones. The overall goal of this di ssertation was to identify “optimal” or “near optimal” experimental designs and statis tical analysis techniques for th e prediction of clonal values and estimation of genetic parameters in or der to achieve maximum genetic gains from clonal testing. In this study, simulations of si ngle-site trials with sets of unrelated clones “planted” in environments with different pa tterns of variabilit y were generated and studied. The main result was that considerable improvement can be obtained through selection of experimental de sign and statistical analysis. In particular, for several experimental designs analyzed with independent errors, the use of single-tree plots (STP) increased the correlations between true and predicted clonal values by 5% over four-tree row plots (4-tree) and yielded greater genetic gain from sele ction. These results indicated that STP allows a more effective sampling of the environmental variation than 4-tree row plots.

PAGE 108

94 Starting with a parametric broad-sense heritability of 0.25 for a completely randomized design, the experimental designs re sulting in the highest heritability increase were row-column (R-C) designs for STP, a nd incomplete blocks with 32 blocks per replication (IB 32) for 4-tree plots. These de signs increased average heritability 10% and 14% over a randomized complete block design, respectively. Unfo rtunately, the best designs always yielded more variable parameter estimates than simpler designs. The use of incomplete blocks (in one or two direc tions) provided explanati on of a larger portion of total phenotypic variability and produced unbiased estimates of genetic variance components and clonal values. For STP, the sm allest incomplete block under study had 8 trees per block and it is possible that sm aller blocks could produce even greater improvements. For the different surfaces patterns simulate d, the ranking from best to worst in terms of easy of accounting for spatial vari ability was ALL, GR AD and PATCH. The latter had the disadvantage th at some of the small patc hes were confounded with the random error; hence, for this surface pattern lower he ritability values should be expected frequently. Also, differences between plot types were smaller for the GRAD surface pattern. Twenty five percent mortality when compar ed to no mortality produced only slight changes in the statistics studied. The conseque nces were mostly reflected in an increase of the variability of some va riance components (variances of clone2ˆ increases about 10%, and variance of 2ˆ about 1%) and, therefore, an incr ease in the variabil ity of individual heritability. Thus, the effect of mortality was small, which could have resulted from the relative large number of ramets used per clone in this study.

PAGE 109

95 For all experimental designs, Dickerson’ s approximate method produced adequate 95% confidence intervals around heritability estimates. They were particularly good for simple designs (completely randomized and ra ndomized complete block), but contained an underestimation in the upper confidence limit for complex designs. The examination of the simulation parameters used to generate the surface patterns showed larger individual heri tabilities when the amount of white noise or unstructured residual (2 ms) was smaller and the spatial correlations larger, with a greater effect of the latter at low levels of 2 ms. Also, the effect of different le vels of gradients on heritability estimates tended to be almost non-existent slightly affecting heritability for GRAD surfaces only. Latinization significantly increased heri tability, but the impact was small in practical terms, and was more important for surfaces with patc hes (ALL and PATCH). However, once an experiment is designed with Latinization, the inclusion in the linear model of its correspond ing effects always yi elded better results. As expected, experiments with more ramets per clone produced more precise variance component estimates a nd larger clonal mean heritab ilities. Using 4 to 6 ramets per clone per site is reco mmended. More than 6 ramets produced only marginal improvements in precision of clonal means. The results from implementing spatial anal ysis techniques indi cated that modeling the error structure other than with independent errors produc ed an important increase in the average correlations between true and pred icted clonal values (CORR), with larger improvements in PATCH surfaces. The incorporat ion of a separable autoregressive error structure with nugget (AR1AR1+) yielded the highest values of CORR. Also,

PAGE 110

96 differences between experimental designs and polynomial models were almost nonexistent when the error structure was modele d with any of the autoregressive options. As expected, the best results were found in the AR1AR1+ error structure with average CORR values as high as 0.90 with the full polynomial regression model, and with average variance components close to their parametric values. Failing to model correctly the data by not in cluding the nugget produced an overestimation of the correlation parameters ( x and y). Comparisons between the best classical de signs that assumed independent errors ( i.e. IB 32 and R-C) in relation to spatial anal yses that used complex error structures, indicated that considerable control for envi ronmental heterogeneity (trend and patches) can be achieved with the traditional classical designs. For example, R-C design had average CORR which were only 0.01 lower than the maximum, and it was the best of the models with independent erro rs for ALL and PATCH surface patterns, and only deficient in GRAD surfaces. Nearest neighbor techniques were also among the best options to account for spatial heterogeneity to increas e individual heritability and ga in from selection. Some of the variants of the Papadakis method were almo st as good as models th at incorporated the error structure, particularly for surfaces with patches. The best models were Papadakis with 4 covariates and no diagonals (PAP-6) a nd a model with 6 cova riates (PAP-11), but due to the simplicity of the model (4 covari ates) the use of PAP-6 is recommended. In this study, the use of an iterated Papada kis method did not produce improvements over the traditional Papadakis.

PAGE 111

97 In summary, the use of spatial analysis te chniques that model the error structure or that use linear covariates are promising options to achieve higher genetic gains, and we recommend their use because they make better use of available information at almost no cost. The last topic was using post-hoc blocking as a way to improve statistical efficiency (or precision) with little effort. The results from post-hoc blocking were promising with negligible differences compared to pre-designed local control. The best post-hoc designs were row-column (for ALL a nd PATCH) and incomplete blocks with 8 blocks (for GRAD). No reduc tion in the genetic variance wa s noted as the size of the block decreased. Also, the use of the log-lik elihood (logL) and the average standard error of the difference (SED) as criterion for se lection of models ar e recommended. While there is no substitute for proper designs of experiments, there are still forestry experiments implemented with randomized co mplete block designs that could benefit from post-hoc blocking. Thus, post-hoc blocki ng should be used wh enever possible and small incomplete blocks are preferred.

PAGE 112

98 APPENDIX A SUPPLEMENTAL TABLES Table A-1. Average variance component estimat es for single-tree plot with no mortality in all surface patterns and experimental designs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design. ALL a TYPE b 2 clone c 2 rep 2 block(rep) 2 col(rep) 2 row(rep) 2 Total CR 0.251 0.751 1.00 RCB 0.252 0.173 0.598 1.02 IB 4 0.250 0.161 0.051 0.560 1.02 IB 8 0.250 0.165 0.069 0.539 1.02 IB 16 0.251 0.169 0.077 0.527 1.02 IB 32 0.250 0.171 0.088 0.513 1.02 R-C 0.252 0.167 0.060 0.047 0.497 1.02 PATCH CR 0.251 0.750 1.00 RCB 0.252 0.023 0.730 1.01 IB 4 0.250 0.016 0.033 0.706 1.00 IB 8 0.251 0.017 0.056 0.681 1.01 IB 16 0.250 0.019 0.071 0.663 1.00 IB 32 0.250 0.020 0.104 0.629 1.00 R-C 0.250 0.016 0.060 0.060 0.618 1.00 GRAD CR 0.252 0.750 1.00 RCB 0.250 0.163 0.608 1.02 IB 4 0.252 0.151 0.047 0.572 1.02 IB 8 0.250 0.157 0.050 0.564 1.02 IB 16 0.250 0.160 0.047 0.563 1.02 IB 32 0.251 0.162 0.045 0.564 1.02 R-C 0.251 0.160 0.040 0.016 0.554 1.02 a ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. b CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. c clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row(rep), row nested within replicate; residual.

PAGE 113

99 Table A-2. Average variance component estimates for four-tree plot with no mortality in all surface patterns and experimental de signs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design. ALL a TYPE b 2 clone c 2 rep 2 block(rep)2 col(rep) 2 row(rep) 2 plot 2 Total CR 0.251 0.250 0.501 1.00 RCB 0.250 0.107 0.197 0.501 1.06 IB 4 0.254 0.083 0.099 0.119 0.501 1.06 IB 8 0.249 0.090 0.140 0.074 0.501 1.05 IB 16 0.249 0.098 0.143 0.063 0.501 1.05 IB 32 0.248 0.102 0.160 0.044 0.500 1.05 R-C 0.252 0.101 0.091 0.029 0.082 0.501 1.06 PATCH CR 0.252 0.140 0.608 1.00 RCB 0.250 0.008 0.140 0.609 1.01 IB 4 0.251 0.006 0.015 0.126 0.608 1.01 IB 8 0.250 0.006 0.026 0.114 0.608 1.01 IB 16 0.251 0.007 0.039 0.102 0.608 1.01 IB 32 0.250 0.007 0.059 0.082 0.607 1.01 R-C 0.250 0.006 0.037 0.024 0.083 0.607 1.01 GRAD CR 0.250 0.192 0.560 1.00 RCB 0.250 0.105 0.140 0.560 1.06 IB 4 0.251 0.079 0.106 0.058 0.560 1.05 IB 8 0.251 0.089 0.133 0.021 0.560 1.05 IB 16 0.251 0.097 0.132 0.015 0.559 1.05 IB 32 0.250 0.101 0.138 0.007 0.558 1.05 R-C 0.251 0.101 0.086 0.015 0.043 0.560 1.05 a ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. b CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. c clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row(rep), row nested within replicate; residual.

PAGE 114

100 Table A-3. Average variance component estimates for single-tree plot with 25% mortality in all surface patterns and experimental designs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design. ALL a TYPE b 2 clone c 2 rep 2 block(rep) 2 col(rep) 2 row(rep) 2 Total CR 0.252 0.751 1.00 RCB 0.253 0.173 0.598 1.02 IB 4 0.251 0.160 0.052 0.559 1.02 IB 8 0.249 0.165 0.070 0.538 1.02 IB 16 0.251 0.168 0.077 0.526 1.02 IB 32 0.250 0.171 0.089 0.512 1.02 R-C 0.252 0.167 0.061 0.048 0.495 1.02 PATCH CR 0.251 0.749 1.00 RCB 0.252 0.023 0.730 1.01 IB 4 0.251 0.017 0.033 0.704 1.00 IB 8 0.251 0.017 0.057 0.680 1.00 IB 16 0.249 0.019 0.073 0.662 1.00 IB 32 0.250 0.021 0.105 0.628 1.00 R-C 0.250 0.017 0.062 0.061 0.615 1.00 GRAD CR 0.252 0.750 1.00 RCB 0.250 0.163 0.607 1.02 IB 4 0.252 0.151 0.047 0.573 1.02 IB 8 0.250 0.157 0.050 0.564 1.02 IB 16 0.250 0.160 0.046 0.563 1.02 IB 32 0.252 0.162 0.045 0.563 1.02 R-C 0.250 0.160 0.040 0.016 0.553 1.02 a ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. b CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. c clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row(rep), row nested within replicate; residual.

PAGE 115

101 Table A-4. Average variance component estimat es for four-tree plot with 25% mortality in all surface patterns and experimental designs simulated. All values are the means for 1,000 simulations and the true values were 2 clone = 0.25 and 2 = 0.75 for CR design. ALL a TYPE b 2 clone c 2 rep 2 block(rep)2 col(rep) 2 row(rep) 2 plot 2 Total CR 0.251 0.250 0.501 1.00 RCB 0.250 0.107 0.197 0.501 1.05 IB 4 0.254 0.082 0.099 0.119 0.502 1.06 IB 8 0.249 0.090 0.141 0.075 0.501 1.05 IB 16 0.249 0.099 0.143 0.063 0.501 1.05 IB 32 0.248 0.102 0.160 0.044 0.499 1.05 R-C 0.252 0.101 0.092 0.030 0.081 0.501 1.06 PATCH CR 0.252 0.142 0.606 1.00 RCB 0.248 0.008 0.142 0.606 1.00 IB 4 0.252 0.006 0.015 0.127 0.607 1.01 IB 8 0.251 0.006 0.027 0.114 0.608 1.01 IB 16 0.250 0.007 0.039 0.101 0.608 1.01 IB 32 0.250 0.007 0.060 0.082 0.606 1.01 R-C 0.250 0.006 0.038 0.025 0.083 0.606 1.01 GRAD CR 0.249 0.192 0.560 1.00 RCB 0.251 0.105 0.139 0.559 1.05 IB 4 0.251 0.079 0.107 0.057 0.561 1.05 IB 8 0.250 0.089 0.134 0.022 0.558 1.05 IB 16 0.250 0.097 0.132 0.017 0.559 1.05 IB 32 0.248 0.101 0.137 0.010 0.557 1.05 R-C 0.249 0.100 0.086 0.015 0.043 0.560 1.05 a ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. b CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. c clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested within replicate; row(rep), row nested within replicate; residual.

PAGE 116

102 Table A-5. Average estimated correlations between true and pred icted clonal (CORR) with standard deviations for 0 a nd 25% mortality obtained from 1,000 simulations for single-tree plots (STP) and four-tree plot row plots (4-tree) in all surface patterns and experimental designs. 0% Mortality STP 4-tree Type a ALL b PATCH GRAD ALL PATCH GRAD CR 0.853 (0.018) 0.852 (0.017) 0.853 (0.017) 0.759 (0.045) 0.796 (0.049) 0.777 (0.032) RCB 0.877 (0.018) 0.856 (0.017) 0.876 (0.016) 0.783 (0.042) 0.799 (0.047) 0.803 (0.029) IB 4 0.882 (0.019) 0.860 (0.018) 0.881 (0.016) 0.816 (0.032) 0.803 (0.044) 0.844 (0.019) IB 8 0.887 (0.019) 0.862 (0.019) 0.882 (0.017) 0.841 (0.028) 0.806 (0.041) 0.865 (0.016) IB 16 0.887 (0.021) 0.862 (0.022) 0.879 (0.017) 0.843 (0.026) 0.810 (0.039) 0.867 (0.016) IB 32 0.887 (0.021) 0.865 (0.021) 0.879 (0.017) 0.849 (0.027) 0.816 (0.040) 0.863 (0.017) R-C 0.888 (0.022) 0.870 (0.024) 0.879 (0.018) 0.821 (0.032) 0.817 (0.039) 0.836 (0.023) 25% Mortality STP 4-tree Type ALL PATCH GRAD ALL PATCH GRAD CR 0.811 (0.023) 0.809 (0.021) 0.810 (0.021) 0.724 (0.043) 0.756 (0.046) 0.739 (0.035) RCB 0.839 (0.023) 0.814 (0.022) 0.838 (0.021) 0.747 (0.040) 0.759 (0.044) 0.764 (0.031) IB 4 0.846 (0.025) 0.819 (0.023) 0.843 (0.022) 0.780 (0.031) 0.763 (0.041) 0.803 (0.024) IB 8 0.850 (0.025) 0.821 (0.024) 0.845 (0.022) 0.803 (0.029) 0.765 (0.039) 0.824 (0.021) IB 16 0.850 (0.026) 0.821 (0.026) 0.842 (0.021) 0.804 (0.027) 0.768 (0.038) 0.823 (0.022) IB 32 0.849 (0.027) 0.824 (0.026) 0.841 (0.022) 0.807 (0.028) 0.774 (0.039) 0.819 (0.022) R-C 0.852 (0.028) 0.829 (0.029) 0.841 (0.022) 0.781 (0.034) 0.776 (0.038) 0.794 (0.027) a CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. b ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches.

PAGE 117

103 Table A-6. Average individual broad-sense heritability calc ulated for single-tree plot (STP) and four-tree plot (4-tree) with their standard deviations in all surface patterns, experimental designs and mortality cases. All values are the means for 1,000 simulations and the base heri tability value was 0.25 for CR design. 0% Mortality STP 4-tree Type a ALL b PATCH GRAD ALL PATCH GRAD CR 0.250 (0.026) 0.251 (0.024) 0.252 (0.025) 0.251 (0.036) 0.252 (0.032) 0.249 (0.032) RCB 0.296 (0.033) 0.256 (0.026) 0.291 (0.030) 0.263 (0.037) 0.250 (0.031) 0.264 (0.033) IB 4 0.309 (0.040) 0.262 (0.030) 0.305 (0.033) 0.290 (0.039) 0.255 (0.032) 0.289 (0.034) IB 8 0.317 (0.045) 0.269 (0.034) 0.307 (0.034) 0.302 (0.040) 0.257 (0.035) 0.302 (0.034) IB 16 0.323 (0.049) 0.273 (0.038) 0.308 (0.035) 0.306 (0.042) 0.261 (0.036) 0.304 (0.034) IB 32 0.328 (0.055) 0.284 (0.044) 0.308 (0.035) 0.313 (0.047) 0.266 (0.039) 0.307 (0.036) R-C 0.336 (0.060) 0.287 (0.052) 0.312 (0.035) 0.301 (0.048) 0.266 (0.041) 0.294 (0.035) 25% Mortality STP 4-tree Type ALL PATCH GRAD ALL PATCH GRAD CR 0.251 (0.029) 0.251 (0.027) 0.251 (0.029) 0.251 (0.041) 0.252 (0.037) 0.249 (0.038) RCB 0.297 (0.035) 0.257 (0.030) 0.292 (0.033) 0.264 (0.042) 0.249 (0.035) 0.265 (0.038) IB 4 0.310 (0.041) 0.263 (0.033) 0.305 (0.036) 0.290 (0.044) 0.255 (0.037) 0.289 (0.039) IB 8 0.317 (0.047) 0.269 (0.037) 0.307 (0.037) 0.301 (0.044) 0.258 (0.038) 0.301 (0.038) IB 16 0.323 (0.052) 0.274 (0.041) 0.308 (0.037) 0.306 (0.046) 0.261 (0.041) 0.303 (0.037) IB 32 0.328 (0.057) 0.285 (0.047) 0.309 (0.038) 0.313 (0.051) 0.266 (0.043) 0.304 (0.040) R-C 0.338 (0.062) 0.289 (0.055) 0.311 (0.038) 0.302 (0.051) 0.266 (0.046) 0.292 (0.038) a CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column. b ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches.

PAGE 118

104 Table A-7. Coefficient of variation (100 std /x) for broad-sense indi vidual heritability obtained from 1,000 simulations with 8 ra mets per clone for single-tree plot in all surface patterns and experimental designs. ALL a Analysis b CR c RCB IB 4 IB 8 IB 16 IB 32 R-C Average No-LAT 10.4 % 11.2 % 12.8 % 14.2 % 15.0 % 16.5 % 17.5 % 13.9 % Partial-LAT 10.3 % 10.9 % 12.9 % 13 .7 % 15.0 % 16.2 % 17.2 % 13.7 % Full-LAT 10.3 % 10.9 % 12.8 % 13.7 % 15.0 % 16.2 % 17.2 % 13.7 % PATCH Analysis CR RCB IB 4 IB 8 IB 16 IB 32 R-C Average No-LAT 9.5 % 10.2 % 11.5 % 12.6 % 13.9 % 15.3 % 17.9 % 13.0 % Partial-LAT 10.0 % 10.4 % 11.3 % 12 .5 % 13.9 % 14.9 % 18.0 % 13.0 % Full-LAT 10.0 % 10.4 % 11.3 % 12.5 % 13.9 % 14.9 % 17.9 % 13.0 % GRAD Analysis CR RCB IB 4 IB 8 IB 16 IB 32 R-C Average No-LAT 9.9 % 10.3 % 10.9 % 11.1 % 11.2 % 11.2 % 11.3 % 10.8 % Partial-LAT 9.6 % 10.5 % 11.0 % 11.2 % 11.3 % 11.7 % 10.9 % 10.9 % Full-LAT 9.6 % 10.5 % 11.0 % 11.2 % 11.3 % 11.6 % 10.9 % 10.9 % a ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches. b No-LAT, design and analysis no La tinized; Partial-LAT, design Latini zed and analysis no Latinized; FullLAT, design and analysis Latinized. c CR, complete randomized; RCB, randomized complete block; IB x incomplete block with x blocks per replicate; R-C, row-column.

PAGE 119

105 Table A-8. Average correlations between tr ue and predicted treatment effects (CORR) with standard deviations in parenthesi s in the ALL surface patterns for 2 and 8 ramets per clone for selected experime ntal designs analys es and polynomial models fitted for the following error structures: independent errors (ID); autoregressive without nugget (AR1AR1); and autoregressive with nugget (AR1AR1+). ID AR1 AR1 AR1 AR1+ ramets / clone ramets / clone ramets / clone Model a 2 8 2 8 2 8 CR 0.674 (0.042) 0.853 (0.018) 0.692 (0.052) 0.886 (0.024) 0.707 (0.061) 0.909 (0.023) RCB 0.674 (0.042) 0.877 (0.017) 0.696 (0.052) 0.891 (0.024) 0.706 (0.060) 0.909 (0.024) R-C 0.695 (0.050) 0.889 (0.022) 0.700 (0.053) 0.893 (0.024) 0.707 (0.059) 0.898 (0.025) Full-Poly 0.692 (0.045) 0.886 (0.019) 0.702 (0.052) 0.894 (0.024) 0.709 (0.055) 0.899 (0.025) a CR, complete randomized; RCB, randomized complete block; R-C, row-column; Full-Poly, full polynomial regression. Table A-9. Original and posthoc blocking average individual broad-sense heritabilities with standard deviations in parenthe sis for all surface patterns obtained from fitting the models in Table 5-1 from simulated datasets with a randomized complete block design with 8 ramets per clone per site. Original Post-hoc Type a ALL b PATCH GRAD ALL PATCH GRAD IB 4 0.309 (0.040) 0.262 (0.030) 0.305 (0.033) 0.310 (0.040) 0.263 (0.030) 0.303 (0.033) IB 8 0.317 (0.045) 0.269 (0.034) 0.307 (0.034) 0.319 (0.045) 0.270 (0.034) 0.307 (0.034) IB 16 0.323 (0.049) 0.273 (0.038) 0.308 (0.035) 0.323 (0.049) 0.275 (0.039) 0.307 (0.035) IB 32 0.328 (0.055) 0.284 (0.044) 0.308 (0.035) 0.329 (0.054) 0.286 (0.043) 0.307 (0.036) R-C 0.336 (0.060) 0.287 (0.052) 0.312 (0.035) 0.336 (0.059) 0.290 (0.053) 0.311 (0.035) a IB x incomplete block with x blocks per replicate; R-C, row-column. b ALL, surfaces with patches and gradients; GRAD, su rfaces with only gradients; PATCH, surfaces with only patches.

PAGE 120

106 APPENDIX B EXTENSIONS IN CONFIDENCE INTERVALS FOR HERITABILITY ESTIMATES A similar study to the one performed in Chapter 4 with STP to obtain 95% confidence intervals was performed with 4tree row plots using the same procedures specified before to compare empirical percen tile with Dickerson’s approximation. In addition, for both plot type s approximated 95% confidence intervals where obtained for analyses from datasets with 25% mortality. The results indicated a surprising agreemen t between the empirical percentiles and Dickerson’s approximations for 4-tree plots, particularly for GRAD surface patterns (Figure B-1). Large differences were detect ed with IB 32 and R-C for surfaces with patches (ALL and PATCH); nevert heless, these differences are irrelevant in practice. In relation to the effects of 25% mortality, as expected, an increase in the ranges of both 95% confidence interval approximation methods was found. This was due to an increase in the variability of the indivi dual heritability estimates (see Table A-6). However, differences detected between empi rical and Dickerson’s approximations for each plot type were consistent for both levels of missing information.

PAGE 121

107 Simulation Dickerson's Figure B-1. Plots of means a nd 95% confidence intervals for broad-sense heritabilities for single-tree plots (STP) and four-tree ro w plots (4-tree) in all designs and surface patterns for non-Latinized designs and analyses (No-LAT) with 25% mortality. The methods correspond to simulation runs and the approximation proposed by Dickerson (1969). Average her itabilities correspond to the central points in each confidence interval. 4-tree Surface: PATCH CRRCBIB 4IB 8IB 16IB 32R-C 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 STP Surface: PATCH CRRCBIB 4IB 8IB 16IB 32R-C 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 4-tree Surface: GRAD CRRCBIB 4IB 8IB 16IB 32R-C 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 STP Surface: ALL CRRCBIB 4IB 8IB 16IB 32R-C 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 STP Surface: GRAD CRRCBIB 4IB 8IB 16IB 32R-C 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 4-tree Surface: ALL CRRCBIB 4IB 8IB 16IB 32R-C 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50

PAGE 122

108 APPENDIX C MATLAB SOURCE CODE Annotated Matlab Code to Generate Error Surfaces function ZST=ar2(n,m,RHOX,RHOY,a,b,s20,e1,e2); % Program that generates the error surfaces % Modified for Exponential = AR(1)xAR(1) (Anisotropic) % Incorporates Polynomial Function for Mean Vector. % Salvador A. Gezan / Dec-2005 % n: number of rows % m: number of columns % RHOX: spatial correlation in the x-direction % RHOY: spatial correlation in the y-direction % a, b: parameters polynomial trend regression % s20: nugget effect error % e1, e2: independent vectors of errors rand('state',sum(100*clock)) s2=1; N=n*m; V=eye(N,N); % Calculation of base V(N*N) MATRIX with AR(1)*AR(1) covariance model Y=zeros(N,1);X=zeros(N,1); for K=1:N Y1=fix((K-1)/m)+1; % I X1=K-(Y1-1)*m; % J Y(K)=Y1;X(K)=X1; % Recording Position for L=K+1:N Y2=fix((L-1)/m)+1; X2=L-(Y2-1)*m; V(K,L)=s2*(RHOX^abs(X2-X1))*(RHOY^abs(Y2-Y1)); % Core AR(1)xAR1(1) V(L,K)=V(K,L); end end % Generation of correlated errors (using Cholesky decomposition) L=chol(V); PAT=L'*e1; PAT=(PAT-mean(PAT))/std(PAT); % Standarized Patches % Incorporation trend and production Surface if a==0 & b==0 Z=PAT; else XC=X-mean(X);YC=Y-mean(Y); mu=a*(XC+YC)+b*((XC.^2).*YC+XC.*(YC.^2)); mu=(mu-mean(mu))/std(mu); % Standarized means (trends) Z=(mu+PAT)./sqrt(2); % Standarized surface end;

PAGE 123

109 % Incorporating Nugget (white noise) ZST=Z*sqrt(1-s20)+e2*sqrt(s20); % Drawing simulated Surface for I=1:n*m; SUR(Y(I),X(I))=ZST(I); end surf(SUR); view(45,30);axis equal; xlabel('X');ylabel('Y'); Annotated Matlab Code to Generate Clonal Trials function CHARS=SIMULA(FORM,NSETS); % Program that generates the clonal trial datasets ready for analysis % Implemented for single site analyses % Salvador A. Gezan / Dec-2005 % Calls files: % SURFALL.TXT Contains 500 surfaces as vectors % Structure: X Y S1 S2 ... S499 S500 % PARAMALL.TXT Contains parameters from surfaces simulated for ALL pattern % PARAMFLAT.TXT Contains parameters from surfaces simulated for ALL pattern % PARAMGRAD.TXT Contains parameters from surfaces simulated for ALL pattern % Structure: SURF RHOX RHOY a b S20 % VALGEN.MAT 500*256 matrix with random clonal genetic values ~N(0,1) % GXENV.MAT 500*256 matrix with random genotype x environment % interactions ~N(0,1) % n: number of rows % m: number of columns % NSETS: number of simulations to be generated % FORM: Plot Type: Single Tree Plot (FORM=1) % 4-tree row Plot (FORM=2) % DSGPTYPE: experimental design type % 1:CR, 2:RCB, 3:IB4, 4:IB8, 5:IB16, 6:IB32, 7:R-C % LATIN: Latinization type % 0:No-Latinized, 1:Latinized % RNDSTP(TYPE,LAT,SET): Function that calls pre-generated CycDesigN for STP % RNDROW(TYPE,SET): Function that calls pre-generated CycDesigN for 4-tree rand('state',sum(100*clock)) % Reading Files SURFS=load('SURFALL.TXT'); % Simulated surfaces SURFS=sortrows(SURFS,[1 2]); PARAM=load('PARAMALL.TXT'); % Parameters of simulated surfaces load VALGEN.MAT % Random vectors of clonal values load GXENV.MAT % Random vectors of gxe interactions % Inicialization n=64; m=32; N=n*m; % Total number of observation CHARS=[]; % Output Matrix SELVG=zeros(500,1); % Selection Vector for Clonal Values SELGXE=zeros(500,1); % Selection Vector for GXE Interaction IDVG=ceil(rand(1)*500); % Random starting vector for clonal values IDGXE=ceil(rand(1)*500);% Random starting vector for gxe interactions

PAGE 124

110 RESP=[]; for SET=1:NSETS % Reads set of Clonal Values IDVG=IDVG+1; if IDVG>500;IDVG=IDVG-500;end BV=VALGEN(:,IDVG); % Reads set of GXE interactions IDGXE=IDGXE+1; if IDGXE>500;IDGXE=IDGXE-500;end GXE=GXENV(:,IDGXE); % Reads SURFACE consecutively ZSUR=SURFS(:,SET+2); X=SURFS(:,1); % x positions in test Y=SURFS(:,2); % y positions in test % Generation of Effects for linear model mu=10; %Population Mean (arbitrary) BV=sqrt(0.20)*BV; % 0.20 s2clone (multiple sites) GXE=sqrt(0.05)*GXE; % 0.05 s2gxe ZSUR=sqrt(0.75)*ZSUR; % 0.75 s2s+s2ms (total error) DSGPTYPE=[1 2 3 4 5 6 7]; for TYPE=1:7 for FORM=1:2 for LATIN=0:1 %Latinization LAT='A'; if LATIN==1;LAT='L';end if FORM==2 & LAT=='L';break;end % Gets pre-generated design (CycDesigN) % Structure NDES:REP BLK PLT TREE LAT CLONE X Y (DSGTYPE 1-6) % Structure NDES:REP ROW COL TREE LROW LCOL CLONE X Y (DSGTYPE 7) if FORM==1 %STP NDES=RNDSTP(TYPE,LAT,SET); elseif FORM==2 %4-tree NDES=RNDROW(TYPE,SET); end % Calculation of Response Values (Yijkl) RESP0=[zeros(size(NDES)) zeros(N,1)]; for I=1:N %Assigns clonal value (VG) and GXE to each CLONE if TYPE==7 VG=BV(NDES(I,7)); GE=GXE(NDES(I,7)); else VG=BV(NDES(I,6)); GE=GXE(NDES(I,6)); end YA=mu+VG+GE+ZSUR(I); % Final response vector % Structure RESP0: REP BLK PLOT TREE LAT (LCOL) CLONE X Y YA % Structure RESP0: REP ROW COL TREE LAT (LCOL) CLONE X Y YA RESP0(I,:)=[NDES(I,:) YA]; % Adds columns with levels of design end

PAGE 125

111 % Saving Final simulated trial file (RESP0) if SET<10 RUN=['00' num2str(SET)]; elseif SET<100 & SET>=10 RUN=['0' num2str(SET)]; else RUN=num2str(SET); end FNAME=[LAT RUN 'T' num2str(TYPE) 'F' num2str(FORM) '.TXT']; if TYPE==7 FFORMAT='%3.0f %3.0f %4.0f %2.0f %3.0f %3.0f … %3.0f %3.0f %3.0f %10.6f \n'; else FFORMAT='%3.0f %3.0f %4.0f %2.0f %3.0f %3.0f … %3.0f %3.0f %10.6f \n'; end fid = fopen(FNAME,'wt'); fprintf(fid,FFORMAT,RESP0'); fclose(fid); end end end % Stores parameters of simulated surface and random vectors used % Structure: SET IDVG IDGXE RHOX RHOY a b S20 CHARS=[CHARS;SET IDVG IDGXE PARAM(SET,2:6)]; end % Saving file of Parameters (CHARS) FNAME='SIMALL.TXT'; FFORMAT='%4.0f %4.0f %4.0f %6.4f %6.4f %6.4f %8.6f %6.4f \n'; fid = fopen(FNAME,'wt'); fprintf(fid,FFORMAT,CHARS'); fclose(fid);

PAGE 126

112 LIST OF REFERENCES Atkinson, A.C. 1969. The use of residuals as a concomitant variab le. Biometrika 56(1): 33-41. Ball, S.T., Mulla, D.J., and Konzak, C.F. 1993. Spatial heterogeneity affects variety trial interpretation. Crop Sci. 33: 931-935. Bartlett, M.S. 1978. Nearest neighbour models in the analysis of field experiments. J. R. Statist. Soc. B 40(2): 147-174. Binns, M.R. 1987. Practical use of neighbor met hods and alternatives in the analysis of field trials. Can. J. Pl ant. Sci. 67: 477-489. Bonga, J.M., and Park, Y.S. 2003. Clonal prop agation, forest tr ees. In: Thomas, B., Murphy, D, and Murray, B. (Eds). Encyclope dia of applied plant sciences. Elsevier Science Ltd. Oxford, U.K. pp. 1395-1402. Box, G.E.P, and Hay, W.A. 1953. A statistical design for the efficient removal of trends occurring in a comparative experiment w ith an application in biological assay. Biometrics 9: 304-319. Brownie, C., Bowman, D.T., and Burton, J. W. 1993. Estimating spatial variation in analysis of data from yield trials: a comparison of methods. Agron. J. 85: 12441253. Brownie. C., and Gumpertz, M.L. 1997. Validity of spatial analyses fo r large field trials. J. Agric. Biol. Environ. Stat. 2(1): 1-23. Carson, M.J. 1986. Advantages of clonal forestry for Pinus radiata – real or imagined? New Zeal. J. For. Sci. 16: 403-415. Cleveland, W.S. 1979. Robust lo cally weighted regression a nd smoothing scatterplots. J. Amer. Statist. Assoc. 74: 829-836. Cochran, W.G. 1957. Analysis of covariance: its nature and uses Biometrics 13: 261281. Cochran, W.G., and Cox, G.M. 1957. Experimental designs. 2nd ed. John Wiley & Sons, Inc. New York. 611pp.

PAGE 127

113 Costa e Silva, J., Dutkowski, G.W., and G ilmour, A.R. 2001. Analysis of early tree height in forest genetic trials is enhanced by including a spatially correlated residual. Can. J. For. Res. 31: 1887-1893. Cressie, N.A.C. 1993. Statisti cs for spatial data – revise d edition. John Wiley & Sons, Inc. New York. 900pp. Cullis, B.R., and Gleeson, A.C. 1991. Spatial analysis of field experiments – an extension to two dimensions. Bi ometrics 47: 1449-1460. Dickerson, G.E. 1969. Techniques for resear ch in quantitative an imal genetics. In: Techniques and procedures in animal science research. Amer. Soc. Anim. Sci. Albany, New York. pp. 36-79. Dieters, M.J. 1994. Inheritance of volume and rust resistance in slash pine. Ph.D. Dissertation, University of Flor ida, Gainesville, Florida. 104pp. Durbn, M., Hackett, C., and Currie, I. 1999. Bl ocks, trend and interfer ence in field trials. In: Friedl, H., Berghold, A., and Kauerm ann, G. (Eds.). Proceedings of 14th International Workshop on Sta tistical Modelling. Graz, Au stria. July 19-23, 1999. pp. 492-495. Dutkowski, G.W., Costa e Silva, J., Gilm our, A.R., and Lopez, G.A. 2002. Spatial analysis methods for forest genetic tr ials. Can. J. For. Res. 32: 2201-2214. Elridge, K., Davidson, J., Harwood, C., and van Wyk, G. 1993. Eucalypt domestication and breeding. Clarendon Press. Oxford, U.K. 288pp. Ericsson, T. 1997. Enhanced heritabilities and best linear unbiased predictors through appropriate blocking of pr ogeny trials. Can. J. Forest. Res. 27: 2097-2101. Falconer, D.S., and Mackay, T.F.C. 1996. Intr oduction to quantitati ve genetics. 4th ed. Prentice Hall. Harlow, U.K. 464pp. Federer, W.T. 1988. Recovery of interblock, in tergradient, and interv ariety information in incomplete block and lattice rectangle de sign experiments. Biometrics 54: 471-481. Fu, Y.B., Clarke, G.P.Y., Namkoong, G., a nd Yanchuk, A.D. 1998. Incomplete block designs for genetic testing: st atistical efficiencies of estimating family means. Can. J. For. Res. 28: 977-986. Fu, Y.B., Yanchuk, A.D., and Namkoong, G. 1999a. Incomplete block designs for genetic testing: some practical consider ations. Can. J. Forest. Res. 29: 1871-1878. Fu, Y.B., Yanchuk, A.D., Namkoong, G., and Clarke, G.P.Y. 1999b. Incomplete block designs for genetic testing: statistical efficiencies w ith missing observations. For. Sci. 45(3): 374-380.

PAGE 128

114 Gilmour, A.R., Cullis, B.R., and Verbyla, A.P. 1997. Accounting for natural and extraneous variation in the analysis of fi eld experiments. J. Ag ric. Biol. Environ. Stat. 2(3): 269-293. Gilmour, A.R., Gogel, B.J., Cullis, B.R., Welham, S.J., and Thompson, R. 2002. ASReml User guide release 1.0. VSN Internati onal Ltd. Hemel Hempstead, HP1 1ES, UK. 267pp. Grondona, M.O., and Cressie, N. 1991. Using spa tial considerations in the analysis of experiments. Technometrics 33(4): 381-392. Grondona, M.O., Crossa, J., Fox, P.N., and Pfeiff er, W.H. 1996. Analys is of variety yield trials using two-dimensional separable ARIMA processes. Biometrics 52: 763-770. John, J.A., and Williams, E.R. 1995. Cyc lic and computer generated designs. Monographs on Statistics a nd Applied Probability 38. Chapman & Hall. London, U.K. 255pp. Johnson, M.E. 1987 Multivariate statistical simulation. John Wiley & Sons, Inc. New York. 230pp. Kempton, R.A., and Howes, C.W. 1981. Th e use of neighbouring plot values in the analysis of variety trials. Appl. Statist. 30(1): 59-70. Libby, W.J. 1977. Rooted cuttings in producti on forests. In: Proc. 14th Southern Forest Tree Improvement Conference, Gainesville, Florida. pp. 13-27. Libby, W.J., and Rauter, R.M. 1984. Advantages of clonal forestry. For. Chron. 60: 145– 149. Lin, C.S., Binns, M.R., Voldeng, H.D., and Guillemette, R. 1993. Performance of randomized block designs in field ex periments. Agron. J. 65: 168-171. Littell, R.C., Milliken, G.A., Stroup, W.W., and Wolfinger, R.D. 1996. SAS system for mixed models. SAS Institute In c. Cary, North Carolina. 633pp. Loo-Dinkins, J., Tauer, C.G., and Lambeth, C.C. 1990. Selec tion system efficiencies for computer simulated progeny te st field designs in loblolly pine. Theor. Appl. Genet. 79: 89-96. Lopez, G.A., Potts, B.M., Dutkowski, G. W., Apiolaza, L.A., and Gelid, P.E. 2002. Genetic variation and inter-trait correlations in Eucalyptus globulus base population trials in Argentina. Forest Genetics 9(3): 217-231. Lui, B.H., Knapp, S.J., and Birkes, D. 1997. Sampling distributions biases, variances, and confidence intervals for genetic corre lations. Theor. Appl. Genet. 94: 8-19.

PAGE 129

115 Magnussen, S. 1993. Bias in genetic variance es timates due to spatial variation. Theor. Appl. Genet. 86: 349-355. MathWorks. 2000. Using MATLAB Version 6.0. The MathWorks, Inc. Natick, Massachusetts. 874pp. Namkoong, G. 1979. Introduction to quantitativ e genetics in forestry. USDA Forest Service Technical Bulletin N 1588. Wa shington, District of Columbia. 342pp. Ndlovu, P. 1992. Classical and bootstrap estima tes of heritability of milk yield in Zimbabwean Holstein cows. J. Diary. Sci. 76: 2013-2024. Patterson, H.D., and Hunter, E.A. 1983. The ef ficiency in incomplete block design in national list and recommended li st cereal variety trials. J. Agric. Sci. 101: 427-433. Patterson, H.D., and Thompson, R. 1971. R ecovery of inter-block information when block sizes are unequal. Biometrika 58: 545-554. Pearce, S.C. 1976. The conduct of “post-mortem s” on concluded field trials. Expl. Agric. 12: 151-162. Pearce, S.C. 1998. Field experimentation on rough land: the me thod of Papadakis reconsidered. J. Agric. Sci. 131:1-11. Pearce, S.C., Clarke, G.M., Dyke, G.V., and Kempson, R.E. 1988. A manual of crop experimentation. Charles Griffin & Company Limited. London, UK. 358pp. Qiao, C.G., Basford, K.E., DeLacy, I.H ., and Cooper, M. 2000. Evaluation of experimental designs and sp atial analysis in wheat br eeding trials. Theor. Appl. Genet. 100: 9-16. Ritchie, G.A. 1992. The commercial use of coni fer rooted cuttings in forestry: a world review. New For. 5: 247-275. Russell, J.H., and Libby, W.J. 1986. Clonal te sting efficiency: the trade-offs between clones tested and ramets per clone. Can. J. For. Res. 16: 925-930. Russel, J.H., and Loo-Dinkins, J.A. 1993. Distri bution of testing effort in cloned genetic tests. Silvae Genetica 42: 98-104. Sarker, A., Singh, M., and Erskine, W. 2001. Effi ciency of spatial methods in yield trials in lentil ( Lens culinaris ssp. culinaris ). J. Agric. Sci. 137: 427-438. Scharf, P.C., and Alley, M.M. 1993. Accounti ng for spatial yield variability in field experiments increases statistical power. Agron. J. 85: 1254-1256. Searle, S.R., Casella, G., and McCulloch, C.E. 1992. Variance components. John Wiley & Sons, Inc. New York. 501pp.

PAGE 130

116 Shaw, D.V., and Hood, J.V. 1985. Maximizing ga in per effort by usi ng clonal replicates in genetic tests. Theor. Appl. Genet. 71: 392-399. Stroup, W.W., Baenziger, P.S., and Mulitze, D.K. 1994. Removing spat ial variation from wheat yield trials: a comparison of methods. Crop Sci. 86: 62-66. Townley-Smith, T.F., and Hurd, E.A. 1973. Us e of moving means in wheat yield trials. Can. J. Plant. Sci. 53: 447-450. Verbyla, A.P., Cullis, B.R., Kenward, M.G ., and Welham, S.J. 1999. The analysis of designed experiments and longitudinal data by using smoothing splines. Appl. Statist. 48(3): 269-311. Vollmann, J., Buerstmayr, H., and Ruckenbaue r, P. 1996. Efficient control of spatial variation in yield trials us ing neighbor plot residuals. Exp. Agric. 32: 185-197. Whitaker, D., Williams, E.R., and John, J.A. 2002. CycDesigN: a package for the computer generation of experimental designs. Version 2.0. CSIRO. Canberra, Australia. 33pp. White, T.L. 2004. Breeding theory and genetic testing. In: Burley, J., Evans, J., and Youngquist, J.A. (Eds). Encyclopedia of fo rest sciences. Elsevier Science Ltd. Oxford, U.K. pp. 1551-1561. Wilkinson, G.N., Eckert, S.R., Hancock, T.W., and Mayo, O. 1983. Nearest neighbour (NN) analysis of field experiments. J. R. Statist. Soc. B 45(2): 151-211. Williams, E.R. 1985. A criterion for the constr uction of optimal neighbor designs. J. R. Statist. Soc. B 47(3): 489-497. Williams, E.R., and Fu, Y.B. 1999. Comment Enhanced heritabilities and best linear unbiased predictors through appropriate blocking of proge ny trials. Can. J. Forest. Res. 29: 1633-1634. Williams, E.R., and John, J.A. 1996. Row-column factorial designs for use in agricultural field trials. Appl. St atist. 45(1): 39-46. Williams, E.R., John, J.A., and Whitaker, D. 1999. Example of block designs for plant and tree breeding trials. Aust. N. Z. J. Stat. 41(3): 277-284. Williams, E.R., Matheson, A.C., and Harw ood, C.E. 2002. Experimental design and analysis for tree improvement. Second Edition. CSIRO Publishing. Collingwood, Australia. 214pp. Wolfinger, R.D. 1996. Heterogeneous vari ance-covariance structures for repeated measures. J. Agric. Biol. Environ. Stat. 1(2): 205-230.

PAGE 131

117 Young, L.J., and Young, J.H. 1998. Statistical ecology: a population pe rspective. Kluwer Academic Publishers. Bost on, Massachusetts. 565pp. Zimmerman, D.L., and Harville, D.A. 1991. A ra ndom field approach to the analysis of field-plot experiments and other spatia l experiments. Biometrics 47: 223-239. Zobel, B.J., and Talbert, J.T. 1984. Applie d forest tree improvement. John Wiley & Sons, Inc. New York. 505pp.

PAGE 132

118 BIOGRAPHICAL SKETCH Salvador Alejandro Gezan was born in October 13, 1973, in Santiago, Chile, a polluted city flanked by the Andes. After a fe w moths, his family moved to Mexico, to settle first in Acapulco and later in Mexico City, where he practically became a Mexican boy. After 10 years he returned to his stil l “unknown” home country to the city of Santiago to live during a few years of political repression and later the happy times of the return to democracy. By 1996 he finished his Bachelor of Science in forestry from the Universidad de Chile in Santiago. A few months later, after being involved in temporary and dubious activities, he obtaine d a research position at the Universidad Austral de Chile in the rainy but pretty city of Valdivia. Hi s job consisted of build ing a growth and yield simulator for native forests, which he enjoye d considerably. After 3 years of mud and wet cold shoes in January of 2001, he started a Ph .D. program in the USA at the University of Florida in the wet, hot and flat city of Gain esville. There, he was lucky to work with two of his most favorite topics, fore stry and statistics; nevertheless at times he was distracted by sports activities such as triathlon. Finally, after 21 years and thanks to the monetary support of several (almost uninterested) sour ces he was able to end his school work.


Permanent Link: http://ufdc.ufl.edu/UFE0012929/00001

Material Information

Title: Optimal Design and Analysis of Clonal Forestry Trials Using Simulated Data
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0012929:00001

Permanent Link: http://ufdc.ufl.edu/UFE0012929/00001

Material Information

Title: Optimal Design and Analysis of Clonal Forestry Trials Using Simulated Data
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0012929:00001


This item has the following downloads:


Full Text












OPTIMAL DESIGN AND ANALYSIS OF CLONAL FORESTRY
TRIALS USING SIMULATED DATA














By

SALVADOR ALEJANDRO GEZAN


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA


2005





























Copyright 2005

by

Salvador Alejandro Gezan




























"If a scientific heresy is ignored or denounced by the general public,
there is a chance it may be right.
If a scientific heresy is emotionally supported by the general public,
it is almost certainly wrong."
(Isaac Asimov, 1977)






Dedicated to:
My biological family, Tita, Lincoln, Ivan, Demain, Florencia and Alexandra
My political family, Dean, Quena and Becky, Lauren, Anha and Kevin
And to an special individual that put us all together, Pincho















ACKNOWLEDGMENTS

I would like to thank the members of my supervisory committee, Drs. T.L. White,

R.C. Littell, D.A. Huber, D. S. Wofford, and R. L. Wu, for their energy, time and help

during my program. I thank, in particular to Dr. White, for the opportunity to come to

Gainesville and do such an interesting project and also for his patience and wiliness to

show me not only science, but also emotional intelligence. I thank to Dr. Huber, for many

small but important details, and for showing me a different dimension of thinking,

sometimes unreachable but which surprisingly I liked. I thank to Dr. Littell, an

unconditional supporter and also model in many senses.

This research would not have been done without the financial support of the

members of the Cooperative Forest Genetic Research Program (CFGRP). Here, I want to

thank Greg Powell for his enormous support in several aspects, and for showing me what

being a gator is all about.

I also want to thank several friends: The Latino Mafia, Veronica, Alex, Rodrigo,

Bernardo, Belkys, Gabriela, and Rossanna; the Trigators Mafia, Terrence, Mathew,

Shannon, Josh, Betsy, Mark, and with special love Eugenia.

Special thanks go to all members of my family for tolerating my absence, but

particularly to my step father Dean W. Pettit, for his unusual vision and for opening doors

for me to cross as I choose.
















TABLE OF CONTENTS



ACKNOW LEDGM ENTS ........................................ iv

LIST OF TABLES ............... ............. ...................... vii

LIST OF FIGURES ................................................. ..............x

ABSTRACT ............................... ............. ...............xiii

CHAPTER

1 IN TR OD U CTION ............... ..1.................... .................. ..........

2 COMPARISON OF EXPERIMENTAL DESIGNS DOR CLONAL FORESTRY
USING SIMULATED DATA....................... .............................7

Introduction ...................... ..... ......................7
M materials an d M eth o d s .......................................................................................... 10
Field Layout. ....... ......... ..... ...... .. ....... ..........10
Genetic Structure and Linear Model .................................. ........12
Spatial Surface..................... ............. ......... 13
Simulation Process ...................... ........ ........ .... ...15
S im u latin g M ortality ...................................................................................... 16
Statistical Analysis ...................... ........ ........ .... ...17
R results and D discussion .............. .................................................................... ............... 19
V ariance C om ponents ...........................................................19
Correlations between True and Predicted Clonal Values...............................22
H eritabilities for Sim ulated D esigns .............................................. ......24
Conclusions....................................... ........ .26

3 ACHIEVING HIGHER HERITABILITIES THROUGH IMPROVED DESIGN
AND ANALYSIS OF CLONAL TRIALS................................................ 29

Introduction ...................................... ......... ........... .29
M materials and M methods ............................................................3 1
Field D esign and Sim ulation .......................................................... ... ....31
Statistical M models ............................. ........................... 33
Results and Discussion ......................... ..... ......... ........ 36
Heritability Estimates and Confidence Intervals...................... ...............36


esign and Sim ulation ......................................................................... 3 1
Statistical M models ........................ ................ ................... ..... .... 33
R results and D discussion ....................... ........ .............. ................................... 36
Heritability Estimates and Confidence Intervals......................................36









Surface P aram eter B behavior ......................................... .......................................40
Latinization........................................ 43
N um ber of R am ets per Clone ........................................................ ... .......... 44
Conclusions and R ecom m endations................................... ...................................... 47

4 ACCOUNTING FOR SPATIAL VARIABILITY IN BREEDING TRIALS ...........49

In tro d u ctio n .............. ...... ...... .. ................. ................. ................ 4 9
M materials and M methods ....................................................................... ..................53
Field D esign and Sim ulation ........................................... ......... ............... 53
Statistical A analysis ........................ .............. ...........................55
Results .............. ........................... ........................59
M modeling Global and Local Trends.................................. ...................... 59
N earest N neighbor M odels......................................................... ............... 65
D discussion .................................... .......... .... .......... ......... ............ 68
M modeling Global and Local Trends.................................. ...................... 69
N earest N neighbor M odels......................................................... ...............72
C o n c lu sio n s........................................................................................................... 7 5

5 POST-HOC BLOCKING TO IMPROVE HERITABILITY AND BREEDING
V A L U E PR E D IC TIO N ................................................................... .....................77

In tro du ctio n ...................................... ................................................ 7 7
M materials and M methods ....................................................................... ..................79
R e su lts ...................................... .......................................................8 3
D iscu ssio n ...................................... ................................................. 8 8

6 CON CLU SION S .................................. .. .......... .. .............93

APPENDIX

A SU PPLEM EN TA L TA B LE S .......................................................... .....................98

B EXTENSIONS IN CONFIDENCE INTERVALS FOR HERITABILITY
E S T IM A T E S ...................................................... ................ 10 6

C M A TLAB SOURCE CODE .......................................................... ............... 108

Annotated Matlab Code to Generate Error Surfaces. ............................................108
Annotated Matlab Code to Generate Clonal Trials. ..............................................109

L IST O F R E FE R E N C E S ....................................................................... .................... 112

BIOGRAPH ICAL SKETCH .........................................................................118
















LIST OF TABLES


Table page

2-1 Simulated designs and their number of replicates, blocks per replicate, plots per
block, and trees per block for single-tree plots (STP) and four-tree row plots (4-
tree). The number of rows and columns is given for the row-column design. All
designs contained 2,048 trees arranged in a rectangular grid of 64 x 32 positions.. 12

2-2 Linear models fitted for simulated datase s for single-tree plots (STP) and four-
tree row plots (4-tree) over all surface patterns, plot and design types. All model
effects were considered random. ........ ..... .......... ........................................ 18

3-1 Linear models fitted for non-Latinized experimental designs over all surface
patterns. All model effects other than the mean were considered random. .........33

3-2 Linear models fitted for Latinized linear experimental designs over all surface
patterns. All model effects other than the mean were considered random. ..........33

3-3 Average individual single-site broad-sense heritability with its standard
deviations in parenthesis and relative efficiencies calculated over completely
randomized design for experimental designs and linear models that were not
Latinized (No-LAT). Note that the parametric HB2 was established for CR
design as 0.25. .....................................................37

3-4 Coefficient of variation (100 std / x) for the estimated variance component for
clone and error on 3 surface patterns, different number of ramets per clone and 3
selected experim ental designs. ........................................................ 45

4-1 Linear models fitted for simulated datasets using classical experimental designs
to model global trends: complete randomized (CR), randomized complete block
(RCB), incomplete block design with 32 blocks (IB), and row-column (R-C)
design. All model effects other than the mean were considered random..............56

4-2 Linear models fitted for simulated datasets using polynomial functions to model
global trends: linear (Linear), quadratic (Red-Poly), and quadratic model with
some interactions (Full-Poly). All variables are assumed fixed and the treatment
effect is random.................................................56


d om .................................................... ................ 56









4-3 Average variance components for the surface pattern GRAD obtained from
fitting the models in Tables 4-1 and 4-2 for the following error structures: a)
independent errors (ID); and b) autoregressive without nugget (AR1 AR1). All
values are the means from 1,000 simulations and the true values were y2g = 0.25
and 2s + 2ms = 0.75 for CR design.................................. ......................... 62

4-4 Average variance components for the surface pattern ALL obtained from fitting
the models in Tables 4-1 and 4-2 for the following error structures: a)
independent errors (ID); b) autoregressive without nugget (AR1OAR1); and c)
autoregressive with nugget (AR1AR1+rj). All values are the means from
1,000 simulations and the true values were 2g = 0.25 and y2s + 02ms = 0.75 for
C R design ............................................................................63

4-5 Average variance components for the surface pattern PATCH obtained from
fitting the models in Tables 4-1 and 4-2 for the following error structures: a)
independent errors (ID); b) autoregressive without nugget (AR1AR1); and c)
autoregressive with nugget (AR1AR1+rj). All values are the means from
1,000 simulations and the true values were 2g = 0.25 and y2s + C2ms = 0.75 for
C R design. ...................................................................64

4-6 Average correlation between true and predicted treatment effects (CORR) and
average treatment variance component for iterated Papadakis methods PAP-6,
PAP-8 and PAP-11 for the surface patterns ALL, PATCH and GRAD .................68

5-1 Linear models fitted for simulated datasets using classical experimental designs
to model global trends over all surface patterns: randomized complete block
(RCB), incomplete block design with x blocks (IB), and a row-column (R-C)
design. All model effects other than the mean were considered random .................81

5-2 Average variance components and individual broad-sense heritability for all
surface patterns obtained from fitting the models in Table 5-1 from simulated
datasets with a randomized complete block designs with 8 ramets per clone per
site ........................................................ .................................85

5-3 Relative frequency of best post-hoc blocking fits from 500 datasets for each
post-hoc design and surface pattern with 8 and 2 ramets per clone per site
according to the log-likelihood (logL) and average standard deviation of the
difference (SED). The values in bold correspond to the most frequent designs......87

A-i Average variance component estimates for single-tree plot with no mortality in
all surface patterns and experimental designs simulated. All values are the
2 2
means for 1,000 simulations and the true values were clone = 0.25 and C2 =
0.75 for C R design. ..................... .................... ..................... .. ......98









A-2 Average variance component estimates for four-tree plot with no mortality in all
surface patterns and experimental designs simulated. All values are the means
2 2
for 1,000 simulations and the true values were clone = 0.25 and 2s = 0.75 for
C R design ............................................................................99

A-3 Average variance component estimates for single-tree plot with 25% mortality
in all surface patterns and experimental designs simulated. All values are the
2 2
means for 1,000 simulations and the true values were clone = 0.25 and C2
0.75 for C R design. ........................................... ........................ 100

A-4 Average variance component estimates for four-tree plot with 25% mortality in
all surface patterns and experimental designs simulated. All values are the
means for 1,000 simulations and the true values were clone = 0.25 and C2s
0.75 for C R design. ........................................... ........................ 10 1

A-5 Average estimated correlations between true and predicted clonal (CORR) with
standard deviations for 0 and 25% mortality obtained from 1,000 simulations for
single-tree plots (STP) and four-tree plot row plots (4-tree) in all surface patterns
and experim ental designs. ........................................ ........................................ 102

A-6 Average individual broad-sense heritability calculated for single-tree plot (STP)
and four-tree plot (4-tree) with their standard deviations in all surface patterns,
experimental designs and mortality cases. All values are the means for 1,000
simulations and the base heritability value was 0.25 for CR design....................103

A-7 Coefficient of variation (100 std / x) for broad-sense individual heritability
obtained from 1,000 simulations with 8 ramets per clone for single-tree plot in
all surface patterns and experimental designs. ................................................ 104

A-8 Average correlations between true and predicted treatment effects (CORR) with
standard deviations in parenthesis in the ALL surface patterns for 2 and 8 ramets
per clone for selected experimental designs analyses and polynomial models
fitted for the following error structures: independent errors (ID); autoregressive
without nugget (AR1OAR1); and autoregressive with nugget (AR1AR1+r). ..105

A-9 Original and post-hoc blocking average individual broad-sense heritabilities
with standard deviations in parenthesis for all surface patterns obtained from
fitting the models in Table 5-1 from simulated datasets with a randomized
complete block design with 8 ramets per clone per site............... ... .................105















LIST OF FIGURES


Figure page

2-1 Replicate layout for randomized complete block design simulations for single-
tree plots (left) and four-tree row plots (right). The same 8 resolvable replicates
were used for incomplete block and row-column designs. All simulations had
256 unrelated clones with 8 ramets per clone for a total of 2,048 trees ................. 11

2-2 Average proportions of restricted maximum likelihood (REML) variance
component estimates (after correction so that all components sum to one) for
single-tree and four-tree row plots for the no mortality case in all surface
patterns and experimental designs simulated. ................. ............................... 21

2-3 Average estimated correlations between true and predicted clonal values
(CORR) obtained from 1,000 simulations for single-tree and four-tree row plots
in the 0% mortality case for all surface patterns and designs. ................................23

2-4 Average heritabilities obtained for single-tree and four-tree row plots in each
surface pattern and design type for the no mortality case. The error bars
correspond to the upper limit for a 95% confidence interval of the mean .............25

2-5 Heritability distributions for the no mortality case in the surface pattern ALL for
selected completely randomized (CR), randomized complete block (RCB), and
row-column (R-C) designs in single-tree and four-tree row plots ..........................25

3-1 Plots of means and 95% confidence intervals for estimated individual broad-
sense heritabilities obtained from simulation runs and by the method proposed
by Dickerson (1969) for all design types and surface patterns for non-Latinized
designs and analyses (No-LAT). Average heritabilities correspond to the central
points in each confidence interval, and all simulations involved 8 ramets per
clon e. .............................................................................. 3 8

3-2 Contours for individual heritabilities obtained by using the lowess smoother
(Cleveland 1979) of row-column design for different surface parameters from
fitting experimental designs and analyses that were non-Latinized (No-LAT). .....39

3-3 Average individual heritability estimates for ALL and PATCH surface pattern
for simulated surfaces for subsets of simulations with High (px > 0.7 and py >
0.7) and Low (px < 0.3 and py < 0.3) spatial correlations............... ................ 42









3-4 Distribution of individual broad-sense heritability estimates for row-column
designs with 2 and 8 ramets per clone obtained from 500 simulations each for
the surface patterns ALL and PATCH from fitting experimental designs and
analyses that were non-Latinized (No-LAT). ................ .................. ...........46

3-5 Average values for row-columns design in all surface patterns from fitting
experimental designs and analyses that were non-Latinized: a) Clonal mean
m heritabilities (the whiskers correspond to the 95% confidence intervals); and b)
Clonal mean heritability increments. ......................................... 46

4-1 Neighbor plots and definitions of covariates used in Papadakis (PAP) and
Moving Average (MA) methods. Plots with the same numbers indicate a
common covariate. ..................... .......... ....... ........ 58

4-2 Average correlations between true and predicted treatment effects (CORR) in 3
different surface patterns for classical experimental design analyses and
polynomial models fitted for the following error structures: independent errors
(ID); autoregressive without nugget (AR1OAR1); and autoregressive with
nugget (AR1AR1+r). Surface GRAD is not shown in the last graph because
few simulations converged. ........... ......... ......................61

4-3 Average correlations between true and predicted treatment effects (CORR) in 3
different surface patterns for nearest neighbor analyses fitted assuming
independent errors. PAP: Papadakis, MA: Moving Average............... ...............67

4-4 Average correlations between true and predicted treatment effects (CORR) in 3
different surface patterns for selected methods: randomized complete block
(RCB), row-column (R-C), and Papadakis (PAP) using the following error
structures: independent errors (ID), and autoregressive with nugget
(AR1OAR1+). ....................................................67

5-1 Experimental layout for randomized complete block design simulations with 8
resolvable replicates (left) and partitioning of one replicate for incomplete block
design layouts (right). All simulations had 256 unrelated clones with 8 ramets
per clone per site for a total of 2,048 trees. .....................................81

5-2 Average correlation between true and predicted clonal values (CORR) for
original and post-hoc blocking analyses on ALL, PATCH and GRAD surface
patterns for simulated surfaces with 8 ramets per clone. .......................................84

5-3 Average individual heritability (a) and correlation between true and predicted
clonal values (b) calculated over 500 simulations with a randomized complete
block designs with 8 ramets per clone per site for each surface pattern and post-
hoc design. ...................................... ......... ........... .86



hoc design ..................................... ........................... .. ..... .. ..... 86









5-4 Average correlation between true and predicted clonal values (CORR)
calculated for ALL and PATCH surface patterns and all post-hoc designs for
sets with High (px, > 0.7 and py > 0.7) and Low (px < 0.3 and py < 0.3) spatial
correlations. The datasets used corresponded to simulated surfaces with a
randomized complete block design and 8 ramets per clone per site ......................89

B-l Plots of means and 95% confidence intervals for broad-sense heritabilities for
single-tree plots (STP) and four-tree row plots (4-tree) in all designs and surface
patterns for non-Latinized designs and analyses (No-LAT) with 25% mortality.
The methods correspond to simulation runs and the approximation proposed by
Dickerson (1969). Average heritabilities correspond to the central points in each
confidence interval. ..................... .. ........................... ...... ... .... ........... 107















Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

OPTIMAL DESIGN AND ANALYSIS OF CLONAL FORESTRY
TRIALS USING SIMULATED DATA

By

Salvador Alejandro Gezan

December 2005

Chair: Timothy L. White
Cochair: Ramon Littell
Major Department: Forest Resources and Conservation

Various alternatives for the design and analysis of clonal field trials in forestry

were studied to identify "optimal" or "near optimal" experimental designs and statistical

techniques for estimating genetic parameters through the use of simulated data for single

site analysis. These simulations investigated the consequences of different plot types

(single-tree or four-tree row), experimental designs, presence of mortality, patterns of

environmental heterogeneity, and number of ramets per clone. Approximated confidence

intervals, Latinization and post-hoc blocking were also studied. Later, spatial techniques

such as nearest neighbor methods and modeling of the error structure by specifying an

autoregressive covariance fitted with two variants (with and without nugget) were

compared.

Considerable improvements were obtained through selection of appropriate

experimental designs and statistical analyses. A 5% higher correlation between true and

predicted clonal values was found for single-tree plots as compared to four-tree plots. The









best experimental designs were row-column for single-tree plots, and incomplete blocks

with 32 blocks for four-tree row plots increasing heritability by 10% and 14% over a

randomized complete block design, respectively. Larger variability of some variance

component estimates was the only effect of 25% mortality. Experiments with more

ramets per clone yielded higher clonal mean heritabilities, and using between 4 and 6

ramets per clone per site is recommended.

Dickerson's approximate method for estimating the variance of heritability

estimates produced reasonable 95% confidence intervals, but an underestimation was

detected in the upper confidence limit of complex designs. The effects of implementing

Latinization were significant for increasing heritability, but small in practical terms. Also,

substantial improvements in statistical efficiency were obtained using post-hoc blocking,

with negligible differences compared to pre-designed local control with no reduction in

the genetic variance as the size of the block decreased.

For spatial analyses, the incorporation of a separable autoregressive error structure

with or without nugget yielded the best results. Differences between experimental designs

were almost non-existent when an error structure was also modeled. Some variants of the

Papadakis method were almost as good as models that incorporated the error structure.

Further, an iterated Papadakis method did not produce improvements over the non-

iterated Papadakis.














CHAPTER 1
INTRODUCTION

Genetic testing is a critical activity for all breeding programs; however, it is time-

consuming and frequently one of the most expensive processes (Zobel and Talbert 1984,

p. 232). Appropriate selection of experimental design and statistical analysis can yield

considerable improvements in terms of increased precision of genetic parameter estimates

and efficient use of resources. Genetic testing aims to achieve several objectives such as

(White 2004) 1) improved genetic gain from selection by evaluation of genetic quality; 2)

estimation of genetic parameters; 3) creation of a base population for future selection

cycles; and 4) quantification of realized genetic gains. Also, various operational decisions

depend on information obtained from genetic tests, and limited quantitative genetic

knowledge degrades the efficiency of a breeding program, eventually affecting potential

gains.

At present, there is extensive research in genetic testing for agricultural crops that

can be useful for testing in forestry; nevertheless, for forest species several important

differences must be considered. Forest sites tend to be heterogeneous, the individuals of

interest (trees) are usually large, and the screening of hundreds or thousands of genetic

entities (e.g., families or clones) is common. These elements imply the need for relatively

large test sites or the use of fewer replicates per site for each entry. Finding optimal

testing sites is difficult, and if areas with high environmental heterogeneity are selected,

the residual error variance will be inflated due to the confounding of tree-to-tree variation

with other large-scale environmental effects. Nevertheless, optimization of genetic testing









is possible in any of the following stages: 1) design of experiments; 2) implementation;

and 3) statistical analysis.

The design of experiments is a critical element and no single design will suit all

testing objectives and environmental conditions perfectly, but statistical and

computational tools are available to improve efficiency and they should be used as

safeguards against site heterogeneity or potential undetected biases. Traditionally, in

forest tree improvement, randomized complete block designs have been the design of

choice, which is effective when within replicate (or block) variability is relatively small,

a situation that is rare in forest sites (Costa e Silva et al. 2001). Nevertheless, more

complex and efficient designs can be easily implemented. Incomplete block designs

allow for a better control of site heterogeneity by specifying smaller compartments than

do randomized complete block design. Also, row and column positions of the

experimental units can be utilized to simultaneously implement two-way blocking factors

producing row-column designs. Other design options include the utilization of restricted

randomization such as Latinization, nested structures and spatial designs (Whitaker et al.

2002).

In the second stage of optimizing experimental precision (i.e., implementation), it is

important to exercise control at all levels of field testing (installation, maintenance and

measurement) with proper care on documentation, labeling, randomization, site

uniformity and survival. The objectives of this stage are to control for all possible factors

that might increase experimental noise, and, therefore, reduce precision of estimation of

parameters of interest.









The stage of statistical analysis is also very important. The use of mixed linear

models combined with techniques such as restricted maximum likelihood (REML) to

estimate variance components and to predict random effects is well understood and

broadly used. Spatial analysis (Gilmour et al. 1997) and nearest neighbor methods

(Vollmann et al. 1996) are interesting techniques used to improve prediction of genetic

values. Spatial analyses are particularly attractive because they incorporate the

coordinates of the experimental units (plots, trees or plants) in the linear model to account

for physical proximity by modeling the error structure (i.e., environmental heterogeneity).

Another relevant tool is post-hoc blocking, which consists of superimposing complete or

incomplete blocks over the original field design and fitting a modified linear model as if

the blocking effects were present in the original design.

Clonal forestry is a new practice in many widely used commercial forestry species,

e.g., Pinus spp., Eucalyptus spp. and Populus spp. (Carson 1986; Elridge et al. 1994, p.

230; Ritchie 1992). Clonal forestry refers to the use of a relatively small number of tested

clones deployed in operational plantations through mass-propagation techniques (Bonga

and Park 2003). Testing clones is of particular interest because the use of identical

genetic entities (ramets from a clone) allows sampling several environmental microsites,

thus allowing complete separation of genetic and environmental effects to increase the

precision of prediction of breeding values (Shaw and Hood 1985). Other benefits of

clonal testing include 1) achieving greater genetic gains than under traditional tree

breeding; 2) capturing greater portions of non-additive genetic variation for deployment;

3) using genotype x environment interaction by selecting clones most suited to specific

site conditions; 4) detecting and utilizing correlation breakers for traits with undesirable









genetic correlations; 5) preventing inbreeding in populations; 6) increasing plantation and

product uniformity; and 7) reducing time between breeding and testing cycles (Libby

1977; Libby and Rauter 1984; Zobel and Talbert 1984, p. 311).

There is extensive research relative to field testing of progenies using seedling from

half or full-sib families (e.g., White 2004). Some of these guidelines can be used for

clonal testing, but differences must be taken into consideration. In the literature, few

studies exist that give important guidelines for the design and analysis of clonal

experiments. The characteristics of the optimal designs and analyses are influenced by

many factors such as magnitudes of heritability, g x e interaction, and ratio of additive to

non-additive genetic variance. Also, several questions of interest remain to be answered.

For example, for the design of experiments: 1) are single-tree plots more efficient than

multiple-tree plots? is there an optimal incomplete block size? 2) how much better than

randomized complete blocks are incomplete blocks or row-column designs? 3) what are

the effects of mortality on the estimation of genetic parameters? 4) do the best designs

depend on the pattern of environmental heterogeneity? and 5) what is the optimal number

of ramets per clone?

For statistical analysis we require answers to the following: 1) is there any gain

from incorporating the error structure in the model? 2) is it necessary to model gradients?

3) how good are the nearest neighbor methods compared with spatial analysis

techniques? and 4) is post-hoc blocking useful; and how should we implement post-hoc

blocking?

One of the best ways to answer many of the above questions is with the use of

simulation techniques. Statistical simulation or Monte Carlo experiments are widely used









in statistical research. These methods are based on the generation of random numbers

with a computer to obtain approximate solutions to problems that are difficult to solve

analytically, and they can aid understanding and knowledge of the properties of the

experiments or methods of interest (Johnson 1987, p. 1). Particularly, for field testing,

they allow several alternatives to be evaluated without incurring large expense or having

to wait years before practical results are available.

The overall goal of this research was to identify "optimal" or "near optimal"

experimental designs and statistical analysis techniques for the prediction of clonal values

and estimation of genetic parameters in order to achieve maximum genetic gains from

clonal testing. This study explored single-site simulations with sets of unrelated clones

"planted" in environments with different patterns of variability.

In the first two chapters many alternative designs for clonal experiments are

compared. In Chapter 2, the consequences for the estimation of genetic parameters of

different design alternatives were studied. The elements compared were 1) single-tree

plot versus four-tree row plots; 2) several experimental designs; 3) no mortality versus

25% mortality; and 4) three different environmental patterns. For Chapter 3, more

detailed work was done with STP experiments. Here, individual heritabilities estimates

were examined in detail according to the different parameters used to simulate the

environmental patterns. Also, confidence intervals for heritability obtained using the

method proposed by Dickerson (1969) were compared with simulated percentile

confidence sets; and the effects of using different number of ramets per clone were

investigated.









Several statistical analysis techniques were studied in the last two chapters using

single-tree plot experiments only. First, in Chapter 4, multiple statistical tools were used

to account for spatial variability. The selected techniques included 1) modeling of global

trends through the use of traditional experimental designs or polynomial models; 2)

specification of a separable autoregressive error structure (with and without nugget); and

3) variants of nearest neighbor methods (Papadakis and Moving Average). Finally,

Chapter 5 aimed to understand consequences of and to define strategies for post-hoc

blocking. Some of the aspects evaluated included comparing performance of original

blocking versus post-hoc blocking for several experimental designs; and defining

strategies to select an optimal blocking structure within post-hoc blocking.














CHAPTER 2
COMPARISON OF EXPERIMENTAL DESIGNS FOR
CLONAL FORESTRY USING SIMULATED DATA

Introduction

Genetic field tests of forest trees are critical to tree improvement serving to

estimate genetic parameters and evaluate provenances, families, and individuals. Genetic

testing of forest trees is a time-consuming process and is frequently the most expensive

activity of an improvement program (Zobel and Talbert 1984, p. 232). Therefore, it is of

primary interest to maximize benefits by allocating resources efficiently (Namkoong

1979, p. 117). It is common in forestry trials to study the performance of a large number

of genetic entries (e.g., families or clones), and this implies the need for relatively large

test sites or the use of fewer replicates per site for each entry. Because forest sites for

field experiments are often inherently variable, it is difficult to find optimal sites. The

presence of this environmental heterogeneity inflates the residual variance due to the

confounding of tree-to-tree variation with other, larger-scale environmental effects; and,

hence, decreases the benefits of using simple experimental designs (Grondona et al.

1996).

Within-site variability is caused by variation in natural factors such as soil,

microclimate, topography, wind and aspect. In addition other forms of heterogeneity

might originate from machinery, stock quality or planting technique. Some conditions

produced by topography or moisture might be easy to identify but, more commonly,

environmental heterogeneity is only recognized after the fact as differential response to









environmental conditions. Gradients across the site, local patches, and random microsite

variance are the common types of variability, and these sources may appear individually

or in combination (Costa e Silva et al. 2001), with some sources more frequent in

particular geographical areas.

There are three stages during which optimization of genetic testing can occur: 1)

experimental design and planning; 2) implementation; and 3) statistical analysis. Forest

genetic trials have traditionally employed randomized complete block designs (RCB) or,

more recently, incomplete block designs (IB). The RCB is most effective when the site is

relatively uniform within replicates, which is rarely the case in forest sites (Costa e Silva

et al. 2001). Reducing the size of the unit by incorporating incomplete blocks within a

full replicate (IB designs) allows for better control of site heterogeneity because smaller

blocks tend to be less variable than larger ones; therefore, IB designs have the potential to

increase precision over RCB (Cochran and Cox 1957, p. 386; Williams et al. 2002, p.

120). For simulated forest genetic trials under several environmental conditions, Fu et al.

(1998) report a 42% increase in the efficiency for IB over RCB. In a related study Fu et

al. (1999a) found that most of the within-site variation can be controlled using 5 to 20

plots per incomplete block with better results in those cases where square blocks were

employed instead of row or column blocks.

Another experimental option is the row-column design (John and Williams 1995, p.

87). When the experimental units are located in two-dimensional arrays, it is possible to

simultaneously implement two blocking factors (instead of one as in IB) corresponding to

the rows and columns of the experiment. Greater efficiencies are expected when row-

column designs are used (Lin et al. 1993; Williams and John 1996). Qiao et al. (2000),









using several wheat breeding trials, reported an increase of efficiency over RCB of 11%

for row-column designs compared with 8% for IB.

Clonal forestry is a new practice in many widely planted commercial tree species,

e.g., Pinus spp., Eucalyptus spp., Populus spp. (Carson 1986; Elridge et al. 1993, p. 230;

Ritchie 1992). Interest in clonal forestry stems from its additional benefits such as ability

to achieve greater genetic gains; potential to capture greater portions of non-additive

genetic variation and genotype x environment interaction; increased plantation and

product uniformity; and accelerated use of results from tree improvement by reducing

breeding and testing cycles (Libby 1977; Zobel and Talbert 1984, p. 311).

In the agronomy and forestry literature, there are several studies that compare

design and analysis alternatives for breeding experiments through real or simulated

datasets. In forest genetics, the few simulation studies available report some guidelines

for future testing. Fu et al. (1998) showed that a-designs were the most efficient

arrangement for IB designs. For the majority of the environmental conditions studied

they were superior to an RCB design for the estimation of family means, but the benefits

of IB over RCB were reduced as the level of missing observations increased (Fu et al.

1999b). In a related study for a-designs, Fu et al. (1999a) reported that smaller

incomplete blocks were more efficient for significant patches when single-tree plots were

established to estimate family and clonal means.

These studies give important guidelines for the design of genetic experiments;

nevertheless, several questions of interest remain to be answered: Are single-tree plots

more efficient than multiple-tree plots? In relation to IB designs, how much better (or

worse) are row-column designs? Is there an optimal incomplete block size? How is the









phenotypic variance partitioned under different experimental designs? Also, it is common

in previous simulation studies to screen a limited number of clones and to treat the

genetic entries as fixed effects in the linear model. In this study genetic entries were

assumed to be random, allowing estimation of heritabilities and prediction of clonal

genetic values for a range of simulated conditions.

The present study is focused on identifying "optimal" or "near optimal"

experimental designs for estimating genetic parameters and achieving maximum genetic

gain from forestry clonal tests. This study explores sets of unrelated clones tested on a

single site through the use of simulated data created with different patterns of

environmental variability. In particular, the objectives of these simulations are to

investigate the consequences for the estimation of genetic parameters of 1) using single-

tree or four-tree row plots; 2) using completely randomized, randomized complete block,

incomplete blocks of various sizes, or row-column designs; 3) experiencing no mortality

versus 25% mortality; and 4) planting on sites with different environmental patterns of

surface variation (only patches, only gradients, and both patches and gradients).

Materials and Methods

Field Layout

All simulations were based on a single trial of 2,048 trees planted on a contiguous

rectangular site of 64 rows and 32 columns with square spacing and 8 ramets for each of

the 256 unrelated clones. The trees were "planted" in either single-tree plots (STP) or

four-tree row plots (4-tree), corresponding to 8 and 2 plots per clone, respectively (Figure

2-1). A completely randomized design (CR) in which clonal plots were randomly

assigned to the field site was considered the baseline experimental design for each plot

size. A randomized complete block design (RCB), a variety of incomplete block designs






11


(IB), and a row-column design (R-C) were also implemented for both STP and 4-tree row

plots (Table 2-1).

There are many ways in which treatments (or clones in this case) can be allocated

to incomplete blocks. For the present work, a-designs were used to obtain IB and R-C

layouts. These designs can be obtained quickly and efficiently for many design

parameters (Williams et al. 1999). All IB and R-C designs were generated using the

software CycDesigN (Whitaker et al. 2002) that incorporates an algorithm to generate

optimal or near-optimal a-designs. Outputs from 100 different independent runs were

used for each design and plot type. CR and RCB layouts were generated using code

programmed in MATLAB (MathWorks 2000).



1 5



2 6 64 trees


3 7

2 32 trees
16 trees 4 8



16 trees 32 trees


Figure 2-1. Replicate layout for randomized complete block design simulations for
single-tree plots (left) and four-tree row plots (right). The same 8 resolvable
replicates were used for incomplete block and row-column designs. All
simulations had 256 unrelated clones with 8 ramets per clone for a total of
2,048 trees.










Table 2-1. Simulated designs and their number of replicates, blocks per replicate, plots
per block, and trees per block for single-tree plots (STP) and four-tree row
plots (4-tree). The number of rows and columns is given for the row-column
design. All designs contained 2,048 trees arranged in a rectangular grid of 64
x 32 positions.

STP
Design a Replicates Blocks/ Rep b Plots/ Block Trees / Block
CR 1 1 2048 2048
RCB 8 1 256 256
IB 4 8 4 64 64
1B 8 8 8 32 32
IB 16 8 16 16 16
IB 32 8 32 8 8
R-C 8 32 rows 32 columns

4-tree
Design Replicates Blocks / Rep Plots / Block Trees / Block
CR 1 1 512 2048
RCB 2 1 256 1024
IB4 2 4 64 256
IB8 2 8 32 128
IB 16 2 16 16 64
IB 32 2 32 8 32
R-C 2 16 rows 16 columns
a CR, complete randomized; RCB, randomized complete block; IB x, incomplete block with x blocks per
replicate; R-C, row-column.
b Rep, resolvable replicate.


Genetic Structure and Linear Model

The linear model used to generate the simulated data was


YVk = + Ck + Es(,jk) + Ems(jk) 2-1

where yijk is the response of the tree located in the ith row and jth column of the kth clone,

LI is a fixed population mean which was set equal to 10 units, Ck is the random genetic

clonal effect, Es(ijk) is the surface error (or structured residual) and Ems(ijk) corresponds to

the microsite random error (or unstructured residual).









The total variance for the above linear model (Equation 2-1) considering all

components as random is C2 T clone + G2s + C2ms. For simplicity, but without loss of

generality, o2T was fixed to 1. Further, the variance structure for all surfaces and designs

was set to c2clone = 0.25 and C2s + c2ms = 0.75; hence, the single-site biased broad-sense

heritability H = o2clone / clone + 2s + 2ms) is 0.25 for completely randomized designs.

Spatial Surface

The spatial surface is a rectangular grid (x and y coordinates) composed of

unstructured and structured residuals. The unstructured errors, Ems, correspond to white

noise and can originate from measurement errors, planting technique, stock quality, and

unstructured microsite variation. The structured residuals, Es, are due to the underlying

environmental surface, and were generated from two distinct patterns: gradients and

patches. Gradients were modeled as a mean response vector t of size 2,048 x 1, at each of

the positions of the 64 x 32 grid, employing the following polynomial function:

t, = a (x, +y,) + p8(x, 2y, + X Y 2) 2-2

where xci and yci correspond to the centered values xci = xi x and yci = yi y for the ith

tree located in column x and row y; and a and representt fixed weights on linear and

quadratic components, respectively. This function defines a flat plane (i.e., no gradients),

when a and flare zero. Also, the primary environmental gradient was oriented along the

short axis of the 64 x 32 rectangle to minimize variation within a replicate, as an

experiment would be laid out in the field.

Patches were modeled by incorporating a covariance structure based on a first-

order separable autoregressive process (AR1OAR1), which is a variant of the exponential

model used in spatial statistics (Littell et al. 1996, p. 305). This error structure has been









previously used successfully for analyses of agricultural experiments (Cullis and Gleeson

1991; Zimmerman and Harville 1991; Grondona et al. 1996; Gilmour et al. 1997) and

forestry trials (Costa e Silva et al. 2001; Dutkowsky et al. 2002). The AR1OAR1 error

structure considers two perpendicular correlations, one for the x direction (px) and the

other for y (py). The model defines an anisotropic model, where the covariance (or

correlation) between two observations is not only a function of their distance, but also of

their direction (Cressie 1993, p. 62). The parameters px and py define the correlation

between the structured residuals (Es) of nearby trees, and there is a positive relationship

between the magnitude of these parameters and patch size.

To generate the patches, a 2,048 x 2,048 variance-covariance matrix (R) was

constructed and later used to simulate residuals. The elements of this matrix were

obtained as

Var(e) = 2s for diagonal elements 2-3


Cov(e,,e,,) = 2s phx Phy for off-diagonal elements 2-4

where hx = I xi xi, 1, hy = I yi yi' |, i.e., the absolute distance in row and column position

respectively, between two trees. The other parameters were previously defined.

This covariance structure also includes a nugget parameter (c2ms), which allows the

modeling of discontinuities in covariance over very small distances. In real datasets the

nugget effect reflects variability that would be found if multiple measurements were

made exactly on the same position or the microsite variation of positions very close

together (Cressie 1993, p. 59; Young and Young 1998, p. 256).









Simulation Process

The first stage of the simulation consisted of the independent generation of surfaces

over which different experimental designs were superimposed. Each surface or

environmental pattern was produced by selecting 5 parameters at random from uniform

distributions (a, Px, p, and C2ms). The parameters were restricted to the following

ranges: 0 to 0.05 for a, 0 to 0.0005 for /, 0.01 to 0.99 for the correlation parameters px

and py, and 0.15 to 0.60 for C2ms. Additionally, these correlations were restricted so their

absolute difference was smaller than 0.85; and C2s was calculated from C2s = 0.75 C2ms.

All these parameters were then used to generate the vector t and the residual

variance-covariance matrix R. The Cholesky decomposition of the R matrix was obtained

and used together with t and an independent random vector of standard normal numbers

to obtain the correlated residuals (Johnson 1987, p. 52-54), producing e ~ N(t, R). Later,

a set of normal independent random residuals was incorporated to constitute the white

noise (i.e., Ems). After this, all components were added together and a standardization was

performed to ensure that the total environmental variance was fixed at 0.75.

The remaining variability (0.25) belongs to the clonal component of the linear

model (Equation 2-1). Because all clones are unrelated, each set of 256 clonal values was

generated as 500 independent standard normal vectors, which were scaled to have a

variance of 0.25.

Three surface patterns were implemented to generate 1,000 independent surfaces of

each pattern: patches only (PATCH) with a= /= 0 in Equation 2-2; gradients only

(GRAD) where px = py = 0 in Equation 2-4; and both patches and gradients together

(ALL) with none of the parameters set to zero. Because of the standardization that was









applied to all surfaces, the comparison between different patterns must be viewed with

caution. ALL includes both gradients and patches and so is more heavily "corrected"

when the error variance was adjusted to 0.75; hence, the effect of each surface component

is reduced when compared with PATCH or GRAD.

The process of generating a simulated trial for a particular surface pattern and

design consisted of selecting at random a surface and a vector of clonal values.

Additionally, a random experimental design (layout) was selected from those generated

through CycDesigN or MATLAB code, and a simple partial randomization was

performed that consisted of shuffling clone numbers. Finally, all components were added

together to produce the response vector (y), which was stored with other relevant

variables (x and y positions, replicate, block, etc.) for statistical analysis.

Simulating Mortality

Two levels of mortality were considered in the present study: 0% and 25%. The

process of mortality was modeled using two independent components: clone and

microsite. The clonal component assumed that some clones survived better than others

because of resistance to disease, adaptability to site, or difference in stock quality. This

component was generated as a random vector of standard normal numbers independent of

their original genetic values (i.e., no relationship with their original clonal values, Ck).

For the microsite component, it was assumed that mortality occurs in clusters or patches;

and it was simulated as a patchy surface with exactly the same error structure previously

described using Equations 2-3 and 2-4 where 20% of the total variation corresponded to

white noise or random mortality. The following values were used to generate the

mortality surfaces: a= /= 0, px = py = 0.75, o2ms = 0.20 and 2s = 0.80. Finally, the










mortality process consisted of calculating an index for each tree based on a weighted

average of both components (10% for clonal and 90% for the microsite component), and

according to this index the lowest 25% of the trees were eliminated. If all ramets (out of

8) were eliminated for any clone, then the process was repeated. The procedure

previously described does not take into account changes in competition due to death of

neighboring trees. This effect was not incorporated because it is expect d to be more

relevant several years after establishment and to affect some variables more than others.

Statistical Analysis

The data for each of the simulated trials were analyzed using ASREML (Gilmour et

al. 2002) with the linear models specified in Table 2-2 (all effects were considered

random). This software fits mixed linear models producing restricted maximum

likelihood (REML) estimates of variance components and best linear unbiased

predictions (BLUP) of the random effect (Patterson and Thompson 1971). Altogether

there were 84,000 datasets analyzed (3 surface patterns x 7 experimental designs x 2 plot

types x 2 levels of mortality x 1,000 simulations each).

The output was compiled and summarized to obtain averages and standard

deviations for heritability estimates and REML variance components for the response

variable y. In addition, empirical correlations (CORR) were calculated between the true

and predicted clonal values.

Single-site heritabilities for each simulation were calculated as

T2 U clone
HB 2 lon+ e for STP, and 2-5
0 clone r0 e

T 2 U clone
HB clone =2 plot for 4-tree. 2-6
2 clone + 02 plot + 2 e







18


Because heritability estimates correspond to a ratio of correlated variance

component estimates, the simple average is not an unbiased estimator of the first

moment. Hence, the following formula was used


for STP, and




for 4-tree


Z ('2 clone + plot + 2j e n


where the sums are over n=1,000 simulations of each surface pattern, plot and design

type.


Table 2-2. Linear models fitted for simulated datasets for single-tree plots (STP) and
four-tree row plots (4-tree) over all surface patterns, plot and design types. All
model effects were considered random.

Design a STP b
CR Yj = p + clone, + r1

RCB y, = + rep, + clone, + ,

IB Yijk = + rep1 + block(rep), + clonek + ijk

R-C Yijk1 = + rep, + col(rep), + row(rep)lk + clones + ijk1

Design 4-tree
CR Yijk = + clone, + plot, + Syk

RCB Yijk = + rep1 + clones + plot, + iyk

IB YijkI = + rep1 + block(rep), + plotik + clonek + Ejk1

R-C Yijklm = + rep1 + col(rep), + row(rep)lk + clone, + plotil + sijklm
a CR, complete randomized; RCB, randomized complete block; IB, incomplete block; R-C, row-column.
b clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate;
col(rep), column nested within replicate; row(rep), row nested within replicate; plot, plot effect; e,
residual.


H (2 cone



2 i clone / l
TT 2


11B









Genetic gain from clonal selection was not estimated in the present study because

gain is directly proportional to heritability (HB2) for a fixed selection intensity and

phenotypic variance (Falconer and Mackay 1996, p. 189). Thus, any increases in HB2 lead

directly to greater genetic gains. Also, the correlation between true and predicted clonal

values (CORR) is a direct measure of genetic gain. As CORR approaches 1, clonal values

are precisely predicted and genetic gain from clonal selection is maximized.

Results and Discussion

Variance Components

For all surface patterns, designs and plot types the simulated data yielded REML

estimates of the genetic variance that averaged close to the parametric values imposed

during simulation (i.e., Y2clone = 0.25). Also, for CR (the reference experimental design)

the estimated variance components for error averaged almost exactly to their parametric

values (i.e., C2s = 0.75 for STP and o2plot + c2 = 0.75 for 4-tree) (Figure 2-2). During the

simulation process, the total variance was set to 1, and, after analysis the estimated

phenotypic variance (sum of average variance component estimates) were extremely

close to 1 for CR for all surface patterns and plot types and for all experimental designs

in PATCH (Tables A-i, A-2, A-3 and A-4).

Simulation scenarios other than CR designs that contained gradients (ALL and

GRAD) and replications resulted in estimated phenotypic variances slightly greater than

1 (2% to 6% overestimation) with higher values for 4-tree row plots than STP. This

situation was due to an inflation of the error variance that originates when a replicate

effect is incorporated in the model. Theoretically, a flat plane is specified for each level

of replicate when in fact a non-horizontal trend should be used to model the field









gradients. The presence of a slightly inflated total variance was corroborated in a small

simulation study using surfaces that contained only gradient and no errors (i.e., random

noise or patches). Upon fitting an RCB to this gradient, residuals were generated and an

error variance was estimated when in fact it should not exist. Box and Hay (1953)

reported a similar situation with time trends and indicated that this inflated error

produced by trends within the replicates can be eliminated by using very small replicates.

In this study, we preferred to correct the variance components in all models so that they

will sum to 1 for each simulated dataset.

For STP in PATCH and ALL surfaces, the variance component for the incomplete

block effect (o2block(rep)) increased with the number of incomplete blocks (Figure 2-2). For

surfaces with patches. the error variance component, C2y, decreased for smaller

incomplete block sizes because an increasing portion of the variance was explained by

incomplete blocks. In general, larger values for the error component Cs2 were found for

PATCH compared to GRAD surfaces indicating that some of the variation due to patches

tended to be confounded with the microsite error instead of being captured by the

incomplete blocks. Fu et al. (1998) reported a similar result where blocks were more

efficient in controlling for gradients than for patches. The variance components 2 rep and

o2block(rep) for those surfaces with gradients (GRAD and ALL) explained a larger portion

of the total variability than for PATCH surfaces. Also, for surfaces with gradients, the

variance components for R-C indicated that there was more variability among columns

than rows, since the main environmental gradient was oriented along the short axis of the

experiment.













4-Tree row Surface: PATCH


CR RCB IB4 IB8 1816 IB 32 R-C


STP -Surface: GRAD 4-Tree row -Surface: GRAD


CR RCB 1B4 1R 16 I 132 R-C


STP-Surface: ALL


CR RCB IB4 B 8 B116 B 32 R-C


4-Tree row- Surface: ALL


CR RCB IB4 I8 1I16 IR 32 R-C


Clone Block(rep) I Row(rep)
I Rep I Col(rep) S Plot Error


Figure 2-2. Average proportions of restricted maximum likelihood (REML) variance
component estimates (after correction so that all components sum to one) for
single-tree and four-tree row plots for the no mortality case in all surface
patterns and experimental designs simulated.


STP-Surface: PATCH









The tendencies for 4-tree were similar to STP except that the values of G2rep were

smaller because replicates were 4 times larger and more heterogeneous. Also, the

proportion of variance explained by incomplete blocks was greater for 4-tree than STP

for GRAD and ALL surfaces and smaller for PATCH. But, the sum of O2rep and o2block(rep)

was very similar between plot types with slightly larger values for STP in PATCH. The

plot effect (o2plot) was smaller for GRAD and larger for PATCH indicating that some of

the variability of patches was confounded in the plot effect. For both plot types, smaller

values of o2, were found on ALL surfaces which had both gradients and patches.

Similarly, Fu et al. (1998) found increased efficiencies in the estimation of family means

when high levels of gradients and patches occurred simultaneously.

The variance components were very similar, comparing 25% mortality to 0%

mortality, for all of the plot types, designs, and surface patterns (Tables A-i, A-2, A-3

and A-4). The major impact of mortality was more variation in the variance component

estimates. For G2clone, the standard deviation increased from 6.8% to 12.6%, while for G2s

the increase was only 0.2% to 2.5%. This result agrees with findings reported by Fu et al.

(1999b) that showed for clonal tests a slight decrease in the efficiency of IB over RCB

under random mortality. However, in the present study the impact of mortality was small

considering that 25% of the observations were missing, and it demonstrates the

usefulness of the REML technique to estimate unbiased parameters efficiently.

Correlations between True and Predicted Clonal Values

Performance of the various designs and plot types for predicting clonal values was

assessed by the correlation between true and predicted clonal values (CORR). For all

surface patterns the best 4-tree row plot design was less precise for predicting clonal











values (lower CORR averages) than any STP design indicating that for these simulations

the latter did a better job accounting for microsite variation (Figure 2-3). This finding

agrees with results reported by Loo-Dinkins et al. (1990) and Costa e Silva et al. (2001).

The difference in average CORR values between the best and worst experimental designs

was larger for 4-tree than for STP, and this difference increased in surfaces that

incorporated gradients (GRAD and ALL). So, while 4-tree row plot designs were

generally less efficient than STP, experimental designs have more impact on the

efficiency of experiments established with multiple-tree plots.

For STP, the best designs were R-C followed closely by IB 32. On GRAD surfaces

all the incomplete block designs behaved similarly, and on PATCH the R-C performed

best. For 4-tree the best designs for GRAD and ALL were IB 8, IB 16 and IB 32; and R-

C was clearly inferior, while for PATCH the best designs were IB 32 and R-C.



STP 0% Mortality 4-Tree row 0% Mortality
090 090

088 088

086 086

084 084 /

082 82

0 80 0 80 *"7'

078 078
-*- PATCH --- PATCH
-0- GRAD / -0- GRAD
076 ALL 076 ALL

074 I I I I- 074
CR RCB IB4 IB8 IB16 IB32 R-C CR RCB IB4 IB8 IB16 IB32 R-C



Figure 2-3. Average estimated correlations between true and predicted clonal values
(CORR) obtained from 1,000 simulations for single-tree and four-tree row
plots in the 0% mortality case for all surface patterns and designs.









The effect of mortality on the average correlations was almost identical for all

designs, plot types, and surface patterns. CORR decreased approximately 4% from the

loss of 25% of the observations (Table A-5).

Heritabilities for Simulated Designs

As for variance components estimates, the average heritability calculated using

Equations 2-7 and 2-8 returned the correct parametric value of HB = 0.25 for the CR

reference experimental design (see Figure 2-4). For the majority of all other experimental

designs, the estimated average heritabilities were always higher than 0.25 indicating that

these designs were effective at reducing the residual variance. For STP, there was a

considerable increase in heritability for RCB above CR in GRAD and ALL surfaces

(Table A-6 and Figure 2-4). For 4-tree row plots, the greater increase occurred when the

design changed from RCB to incomplete block designs. In STP, the design producing the

highest average HR for any surface pattern was R-C followed by IB 32. In the case of 4-

tree, the best design was clearly IB 32.

For the IB designs, larger heritability values were obtained as the number of

incomplete blocks increased for all plot types and surface patterns indicating that smaller

incomplete blocks were more efficient than larger blocks in controlling environmental

variation. Similar conclusions were obtained from other simulation studies where

between 5 and 10 plots (or trees) per incomplete block were recommended (Fu et al.

1999a). For this study, the best incomplete block design (IB 32) had only 8 plots per

block for both STP and 4-tree row plots. This block size is relatively small, and the use of

fewer plots could produce greater heritabilities, but further studies are required.














4-Tree row 0% Mortality


CR RCB IB4 IB8 IB 16 IB32 R-C


CR RCB IB4 IB8 IB 16 IB32 R-C


Figure 2-4. Average heritabilities obtained for single-tree and four-tree row plots in each
surface pattern and design type for the no mortality case. The error bars
correspond to the upper limit for a 95% confidence interval of the mean.


STP 0% Mortality


4-Tree row 0% Mortality


150


100


50 / so -50




0.1 2 0.3 0.4 0.5 0.6 0.1 0.2 0.3 0.4 0.5 0.6
HB2 H,2

Figure 2-5. Heritability distributions for the no mortality case in the surface pattern ALL
for selected completely randomized (CR), randomized complete block (RCB),
and row-column (R-C) designs in single-tree and four-tree row plots.


036


034


032


030


M 028
I
026


024


022


020


STP 0% Mortality


034


032

030
I,
M 028
I
026


024


022

020


STP 0% Mortality









The comparison of the distributions of H2 estimates for different designs showed

that as the number of incomplete blocks increased the distribution became asymmetric,

with a longer tail and increased spread to the right (Figure 2-5). This result indicates that

assuming a normal distribution for heritability can sometimes be incorrect.

A z-test comparing the H, values for 0% and 25% mortality levels for designs of

the same plot type and surface patterns was conducted to study if the heritability

estimates changed. No statistically significant differences were found (experiment-wise a

= 0.05) among average HB2 estimates with the exception of a few comparisons within R-

C indicating that the analyses produced unbiased heritability estimates for all

experimental designs and plot types. The main effect of mortality was the increase in the

variance among HR estimates. Compared to 0% mortality, Var(HB2) increased 9.8% for

STP and 12.0% for 4-tree row plots for datasets with 25% missing trees.

The minor effect produced by 25% missing values could indicate that 8 ramets per

clone was more than necessary. Several authors recommend using between 1 and 6

ramets/clone for single site experiments (Shaw and Hood 1985, Russell and Libby 1986);

hence, it is possible that under the conditions of the present study fewer than 8 ramets per

clone might be adequate; nevertheless, this topic requires further study.

Conclusions

The results from these simulations indicate that proper selection of experimental

designs can lead to considerable increases in heritability, precision of predicted genetic

values and genetic gain from selection. The use of single-tree plots instead of 4-tree row

plots resulted in an average increase in the correlation between true and predicted clonal

values of 5%. Hence, single-tree plots allow a more effective sampling of the









environmental variation and reduce error variance more than 4-tree row plots. The best

experimental design for single-tree plot experiments was the row-column design which

fits random effects for both rows and columns within each resolvable replicate. R-C

designs were followed closely by incomplete block designs with small blocks which were

very efficient for STP. For 4-tree row plots, an incomplete block design with 32 blocks

per replicate is recommended.

The use of incomplete blocks (in one or two directions) controlled for an important

portion of the total environmental variability and produced unbiased estimates of genetic

variance components and clonal values. The increase in the standard deviation of the

average pair-wise clonal comparison from the use of incomplete blocks was counteracted

by the improved precision obtained in the statistical analysis. For single-tree plots, the

smallest incomplete block under study had 8 trees per block, and it is possible that

smaller blocks could produce even greater improvements.

For the different simulated surface patterns, the ranking using heritability or

correlation from high to low was ALL, GRAD and PATCH. In the latter pattern some of

the small patches were confounded with random error; hence, for this surface pattern

lower heritability values should be expected.

Twenty-five percent mortality produced only slight changes in the statistics studied.

The consequences were primarily an increase in the variability of some variance

components (variance of 2clione increases about 10%, and a2s about 1%) and, therefore,

the variability of heritabilities increased. Thus, the effect of mortality was small, but

might have been ameliorated by the relatively large number of ramets used per clone.






28


Lastly, the simulations from this study did not incorporate the effect of competition

between neighboring trees; therefore, these results must be interpreted with caution.

Trials with strong between tree competition, as occurs in older testing ages, and with

large mortality patches, might produce different results.














CHAPTER 3
ACHIEVING HIGHER HERITABILITIES THROUGH IMPROVED
DESIGN AND ANALYSIS OF CLONAL TRIALS

Introduction

For any operational breeding program, genetic testing constitutes one of the most

important and expensive activities. Several alternatives are available for experimental

designs. The widely employed randomized complete block design (RCB) is most

effective when blocks (or replicates) are relatively uniform, something that usually occurs

only with small replicates (Costa e Silva et al. 2001). With large numbers of treatments

(e.g., families and clones) the use of incomplete blocks (IB) can increase efficiency

considerably (Fu et al. 1998; Fu et al. 1999a), particularly when there are large amounts

of environmental variability or when the orientation of replicates can not be correctly

specified (Lin et al. 1993). Also, the row and column positions of the experimental units

can be utilized to simultaneously implement two blocking factors and produce row-

column designs (R-C), as described in detail by John and Williams (1995, p. 87). These

designs have demonstrated greater efficiencies than other common designs (Lin et al.

1993; Williams and John 1996). Latinization is rarely used as a design technique for

increasing the efficiency of estimating treatment effects. A design is said to be

"Latinized" when the randomization is restricted so that the position of the experimental

units for the same treatment are forced to sample different areas of the experimental area.

This is accomplished by defining long blocks (such as row or columns) that span multiple









replicates ensuring that the treatments are spread out across the entire test site (John and

Williams 1995, p. 87-88).

The increased interest in clonal forestry is generated by the higher genetic gains

that can be obtained from using tested clones for deployment to operational plantations.

Some benefits include the possibility of capturing non-additive genetic effects, using

greater amounts of genotype x environment interaction, and increasing plantation and

product uniformity (Zobel and Talbert 1984, p. 311).

Field experiments for clonal testing are particularly challenging because large

numbers of genotypes need to be evaluated, implying the need for large test areas.

Relatively homogenous areas are difficult to find because forest sites tend to have high

environmental variability usually expressed in the form of patches, gradients or both,

together with considerable random microsite noise (Costa e Silva et al. 2001). Therefore,

site heterogeneity is an important factor to consider when clonal trials are designed,

implemented and analyzed. Selection of an appropriate experimental design together with

a correct specification of the linear model could produce considerable improvements in

the precision of predicted genetic values, heritabilities and gains from selection.

Several recommendations are available in the literature for clonal testing, and in

general, optimal designs would employ 1 to 6 ramets per clone per site, and the number

of clones tested would be maximized while utilizing as few ramets as possible (Shaw and

Hood 1985; Russell and Libby 1986; Loo-Dinkins et al. 1990; Russell and Loo-Dinkins

1993). Still, difficulties remain in defining optimal designs, numbers of families, clones

per family and ramets per clone to be planted on single or multiple sites. Also, it is not

clear which analytical methods are to be preferred.









In Chapter 2, we simulated single-site clonal trials under several experimental

designs and 3 different environmental surface patterns to identify appropriate conditions

for estimating genetic parameters. That study showed that single-tree plots (STP)

experiments were more efficient for predicting clonal values than four-tree row plots

across all designs and conditions simulated. This new study considers only STP

experiments and aims to determine the effects of different designs and analytical options

on the prediction of clonal values and the precision of some genetic components. In

particular, the objectives are to determine 1) which experimental designs, including

Latinization, maximize broad-sense heritability and, hence, gain from clonal selection? 2)

which patterns of environmental or spatial variability yield high or low heritabilities? 3)

what the effects of using different number of ramets per clone are? and 4) how close

Dickerson's approximate method for confidence intervals is to the empirical estimates?

Materials and Methods

Field Design and Simulation

The simulated datasets used in this study are based on a single site trial of 2,048

ramets planted in a rectangular grid of 64 rows and 32 columns with 8 ramets for each of

the 256 clones arranged in single-tree plots with no missing observations (details in

Chapter 2). Briefly, two factors were considered in simulating the environments under

which the different experiments were performed a gradient generated with a polynomial

function depending on the x and y coordinates of the grid; and patches that were modeled

by incorporating a covariance structure based in a first-order separable autoregressive

process or AR1OAR1 (Cressie 1993; Gilmour et al. 1997). The polynomial function

employed was









t = a(x Y) + P+(xc 2y +xl 2) 3-1

where xc and yci correspond to the centered values xc = xi x and yci = yi y for the ith

tree located in column x and row y; a and representt fixed weights on linear and

quadratic components, respectively.

The AR1OAR1 error surface employed two perpendicular correlations (px and p,)

and was generated through an error variance-covariance matrix defined as

Var (e) = cs + cms for diagonal elements 3-2

Cov (e,, e,) = sphxpyhy for off-diagonal elements 3-3

where hx = I xi xi' 1, hy = I yi yi' 1, i.e., the absolute distance between two trees in the

row and column position, respectively. The variance component c2s describes the surface

error or structured residual, and 2yms corresponds to the microsite random error, white

noise or unstructured residual. The three surface patterns simulated included only patches

(PATCH), only gradients (GRAD), and a combination of both (ALL). Each surface was

generated by drawing at random the parameters a, /, px, py and 2yms from uniform

distributions with the following ranges: 0 to 0.05 for a, 0 to 0.0005 for /, 0.01 to 0.99 for

px and py, and 0.15 to 0.60 for 2Ms. Without loss of generality, the total variance for all

simulated datasets was fixed at 1, and the genetic structure was set to have a single site

biased broad-sense heritability of 0.25; hence, G2clone = 0.25 and C2y + C02m = 0.75.

The experimental designs studied were completely randomized (CR), randomized

complete block with 8 resolvable replicates (RCB), a variety of incomplete block designs

with 4, 8, 16 and 32 incomplete blocks within each resolvable replicate (IB 4, IB 8, IB 16

and IB 32, respectively), and a row-column design (R-C). Both non-Latinized and

Latinized designs were obtained for IB and R-C designs using the software CycDesigN











(Whitaker et al. 2002). One thousand simulations were generated for each combination of

surface pattern, design type and Latinization option, for a total of 36,000 datasets.

Statistical Models

All the simulated experiments were analyzed using the software ASREML

(Gilmour et al. 2002) fitting the models specified in Tables 3-1 and 3-2 for non-Latinized

and Latinized design options, respectively. For each set of simulations, the restricted

maximum likelihood (REML) variance components estimates were obtained, and single-

site individual heritabilities for each fitted model were estimated using the following

expression:


H"=- "- 3-4
2 clone 3 e
H -2cone + j2e


Table 3-1. Linear models fitted for non-Latinized experimental designs over all surface
patterns. All model effects other than the mean were considered random.

Design a Model b
CR yj = p + clone, + r1
RCB y, = p + rep, + clone, + r,
IB Yijk = i + rep + block(rep), + clonek + Eik
R-C Yijk1 = i + rep + col(rep), + row(rep),k + clones + Eijkl
a CR, complete randomized; RCB, randomized complete block; IB, incomplete block; R-C, row-column.
b clone, clone effect; rep, resolvable replicate; block(rep), incomplete block nested within replicate;
col(rep), column nested within replicate; row(rep), row nested within replicate; e, residual.



Table 3-2. Linear models fitted for Latinized linear experimental designs over all surface
patterns. All model effects other than the mean were considered random.

Design a Model b
IB Yijkl = 1i + rep, + lat, + block(rep)lk + clone1 + Eljk
R-C Yijklmn = .t + repi + lcolj + lrOWk + col(rep),i + row(rep)lm + clone + Sijklmn
a IB, incomplete block; R-C, row-column.
b clone, clone effect; rep, resolvable replicate; lat, Latinized block; block(rep), incomplete block nested
within replicate; Icol, Latinized long column; Irow, Latinized long row; col(rep), short column nested
within replicate; row(rep), short row nested within replicate; e, residual.









To estimate an unbiased empirical expected heritability for each design and surface,

the ratio of the mean of the numerator over the mean of the denominator was used (see

Equation 3-4). Also, the relative efficiency (RE) was calculated as the ratio of means of

the heritability of a particular design divided by that of an RCB design of the same

surface pattern.

In this study, Dickerson's method was used to obtain approximate 95% confidence

intervals (CI's) on individual heritability estimates. Approximations are required because

heritabilities correspond to a ratio between linear functions of random variables, and in

general, and particularly for unbalanced data, a closed form expression does not exist to

estimate confidence intervals (Dieters 1994, p. 4). Assumptions and procedures

suggested by Dickerson (1969) approximate the variance of H B for each data set as


(f 2 )= Vr (j 2 clone) 3-5
Var(H ) = r one) 3-5
k[ clone + 2 e]2

where the numerator Var(j2 clone) corresponded to the variance of a variance component

obtained from ASREML based on asymptotic theory of restricted maximum likelihood

estimation (Searle et al. 1992, p. 473). Then, Dickerson's 95% confidence intervals for

Hi2 were calculated as HE2 1.96 x std(H2) where std(HB2) is the square root of the

variance expression from Equation 3-5. For comparison, empirical or percentile CI's

were obtained by finding the top and bottom 2.5% estimated Hf 2 from the 1,000

individual heritability estimates from each fitted set of simulations. CI's were calculated

for each experimental design and environmental surface.

The magnitude of fj 2 is affected by different levels of the parameters used to

generate the error surfaces (a, ,/, px, py and Cms). Influences of these parameters were










studied using heritability estimates obtained from the non-Latinized designs and models

fitted without Latinization through the use of simple group means and by producing

smoothed three-dimensional surfaces using the loess smoother (Cleveland 1979).

Two types of comparisons were made to study the efficiency of Latinization. First,

the effectiveness of Latinization in the design phase alone was tested using a z-test with

the estimated heritabilities that were generated with and without Latinization analyzed

with linear models that did not include these Latinization effects (i.e., only models from

Table 3-1). These analyses correspond to Partial-LAT (indicating Latinization in design

but not in analysis) and No-LAT (no Latinization for both design and analysis). Second,

to examine the importance of including Latinization in the analysis phase, Latinized

experimental designs were fitted with and without Latinization effects in their respective

linear models. These analyses were identified as Full-LAT (indicating Latinization in

design and analysis) and Partial-LAT, and the likelihood ratio test for mixed models

described by Wolfinger (1996) was used to compare these fitted models.

Finally, the effects of using different numbers of ramets per clone on the efficiency

of estimating variance components and predicting clonal values were studied by selecting

subsets of each original simulated dataset and fitting their corresponding linear models.

The first 500 simulations of three designs (RCB, IB 32 and R-C) previously generated for

the non-Latinized designs were analyzed by selecting at random 2, 4, 6 and 8 contiguous

replicates and using the linear models from Table 3-1. Comparisons were preformed

using both individual broad sense heritability and mean clonal heritability (H,-2) from

each subset. The Hfi2 formula used was

'2
T-2 clone 3-6
cc 3-6
0 clone +T e I 1m









where the variance components are as previously described, and m is the number of

ramets (2, 4, 6 or 8). As with Equation, 3-4 an unbiased mean was estimated as the ratio

of the mean of the numerator over the mean of the denominator.

In the present study comparisons of methods and designs were made using

heritabilities only. Genetic gains are also of interest, but they were not used here because

for a given intensity of selection and fixed phenotypic variance, gains are directly

proportional to heritability (Falconer and Mackay 1996, p. 189).

Results and Discussion

Heritability Estimates and Confidence Intervals

The R-C design produced the highest average individual heritability and relative

efficiency (RE) for all surface patterns (Table 3-3 and Figure 2-4). This was particularly

true for surface pattern ALL where both gradients and patches were simulated together.

For IB designs, heritability increased as the size of the incomplete block decreased for all

surfaces, and IB designs with 32 incomplete blocks per replicate (8 trees per block) were

nearly as efficient as R-C designs for all surfaces. In general, for all designs higher

relative efficiencies were found for surfaces with simulated patches (ALL and PATCH),

indicating the usefulness of one or two-way blocking to account for them. These results

agree with other studies that also reported greater efficiencies for IB and R-C designs

(Lin et al. 1993; Fu et al. 1999a; Qiao et al. 2000). Also, higher average heritabilities

were accompanied by an increase in their standard deviation (Table 3-3 and Figure 2-5).

For example, in PATCH surfaces for R-C designs, the standard deviation was more than

double that of RCB designs. Hence, the implementation of genetic experiments with R-C

and IB 32 designs is recommended; nevertheless, they have the drawback of yielding

more variable heritability estimates.










Table 3-3. Average individual single-site broad-sense heritability with its standard
deviations in parenthesis and relative efficiencies calculated over completely
randomized design for experimental designs and linear models that were not
2
Latinized (No-LAT). Note that the parametric H2 was established for CR
design as 0.25.

HE2 RE

Design a ALL b GRAD PATCH ALL GRAD PATCH
0.250 0.252 0.251
CR 0.84 0.86 0.98
(0.026) (0.025) (0.024)
0.296 0.291 0.256
RCB 1.00 1.00 1.00
(0.033) (0.030) (0.026)
0.309 0.305 0.262
IB 4 309 0305 1.04 1.05 1.02
(0.040) (0.033) (0.030)
0.317 0.307 0.269
IB 8 1.07 1.05 1.05
(0.045) (0.034) (0.034)
0.323 0.308 0.273
IB 16 1.09 1.06 1.07
(0.049) (0.035) (0.038)
0.328 0.308 0.284
IB 32 1.11 1.06 1.11
(0.055) (0.035) (0.044)
RC 0.336 0.312 0.287
R-C 1.13 1.07 1.12
(0.060) (0.035) (0.052)
a CR, complete randomized; RCB, randomized complete block; IB x, incomplete block with x blocks per
replicate; R-C, row-column.
b ALL, surfaces with patches and gradients; GRAD, surfaces with only gradients; PATCH, surfaces with
only patches.


In most cases, Dickerson's approximation and the percentile CI's obtained directly

from simulations gave similar estimates for the 95% confidence intervals (Figure 3-1).

Both methods were particularly close for the RCB designs for all surface patterns, and for

all designs in the GRAD surfaces. In contrast, Dickerson's approximation almost always

yielded smaller CI's than the percentile intervals for surfaces PATCH and ALL, and the

upper limit of the CI's was particularly underestimated for R-C and IB designs on these

surfaces. This was a consequence of Dickerson's approximation assuming a symmetrical

distribution when, in fact, the simulated H 2 distributions were skewed to the right (see

Figures 2-5 and 3-4).










Surface: PATCH


0 25

0 20

0 15


0 50

0 45

0 40

0 35

0 30

0 25

0 20

0 15


0 50

0 45

0 40


CR RCB IB 4 IB 8 IB 16 IB 32 R-C
Surface: GRAD









S00 0l



CR RCB IB 4 IB 8 IB 16 IB 32 R-C
Surface: ALL





jilt


CR RCB IB 4 IB 8 IB 16 IB 32


R-C


-0- Simulation -4- Dickerson's


Figure 3-1. Plots of means and 95% confidence intervals for estimated individual broad-
sense heritabilities obtained from simulation runs and by the method proposed
by Dickerson (1969) for all design types and surface patterns for non-
Latinized designs and analyses (No-LAT). Average heritabilities correspond
to the central points in each confidence interval, and all simulations involved
8 ramets per clone.


: Iii


-

-





-

-
















H2- Surface PATCH


00 02 04 06 08 10 00 02 04 06 08 10
PX Px

H2-Surface GRAD H2-Surface GRAD
00005 0 8


001 002 003

H2 Surface ALL


001 0 02 0 03 0 04 0 05

H2 -Surface ALL


00 02 04 06
PX


I 02
08 10 00 02



0 25 0 30 0 35 0 40 0 45 0 50


04 06 08 10


Figure 3-2. Contours for individual heritabilities obtained by using the lowess smoother

(Cleveland 1979) of row-column design for different surface parameters from

fitting experimental designs and analyses that were non-Latinized (No-LAT).


0 0004 -




0 0003-




0 0002-




0 0001 -


0 0000 ---
0 00


H2-Surface PATCH









The ideal situation would be to derive the exact probability distribution for

heritability, but this is very complex and usually only possible for simple linear models

and balanced datasets. Better approximate confidence intervals than Dickerson's can be

obtained using Taylor series, which improves the estimation of the heritability variance

(more details in Dieters 1994, p. 17-20); nevertheless, methods based on Taylor series do

not correct the problem of skewness. Another option is to use bootstrap and jackknife

techniques to approximate distributions by re-sampling. These methods have been

implemented successfully for specific situations with heritability (Ndlovu 1992) and

genetic correlations estimates (Liu et al. 1997), but their statistical properties are not well

understood.

Surface Parameter Behavior

The simulation parameters (a, px, py and C2ms) used to generate the three surface

patterns varied considerably and differentially influenced the magnitude of the individual

heritability estimates. Gradients (or trends) on simulated surfaces were specified by a and

,8, here, larger values corresponded to steeper gradients. These parameters, in most cases,

had little impact on heritability estimates; with larger values of a and /, only small

increases in heritabilities were found. Also, as expected, the effect of these gradient

parameters was more pronounced in GRAD surfaces (for R-C see Figures 3-2c and 3-2d).

It is well known that smaller amounts of unstructured residual increase the

efficiency of analyses of field experiments. For this study, an increase in heritabilities for

all designs and surface patterns was noted as the simulated 02ms decreased. This increase

was more pronounced for designs with smaller incomplete blocks, where some










2 2
heritability estimates reached values as high as 0.50. Also, increases in H,2 as the 2ms

got smaller were more pronounced in surfaces with patches (ALL and PATCH).

For the simulated surfaces, using larger spatial correlation parameters (px and py)

produced bigger patches. The results of this study indicated that, in general, for all

experimental designs there was an improvement in heritabilities as the patch size

increased. The effect of the spatial correlation parameters was larger for IB and R-C

designs than for RCB designs. Also, an interaction between these parameters and &2ms

was observed. For example, in R-C designs, high values of spatial correlation and low

values of white noise yielded larger heritabilities, where the effect of spatial correlations

was more pronounced at low values of O2ms (Figures 3-2b and 3-2f).

In another simulation study Fu et al. (1998) reported that increasing patch size in

RCB and IB designs gave smaller variance of family mean contrasts, a result that agrees

with our findings. Nevertheless, they also reported that the presence of gradients

produced a reduction in precision, which was more pronounced for RCB designs. In the

present study, the effect of gradient was almost negligible, but more relevant for those

designs that had smaller incomplete blocks. The disparity of results is probably a

consequence of using different block sizes. Fu et al. (1998) considered a long rectangular

block (10 x 1 tress) perpendicular to the main gradient; therefore, increasing its effect, but

in the present study blocks were always considered square or almost square.

The larger effect of px and py on IB and R-C designs in comparison to RCB occurs

because small incomplete blocks were more likely to be completely contained in a larger

patch; therefore, reducing the within block variability and explaining a larger portion of

the total variance in comparison to using only large blocks or replicates. To examine this










fact, the average individual heritabilities for subsets of simulations with low spatial

correlations (px < 0.3 and py < 0.3) and large correlations (px, > 0.7 and py > 0.7) were

compared. For all surface patterns, within the set of low correlations, no important

differences were noted between all designs, but in the case of high correlations, an

important increase in average heritability occurred as the size of the incomplete block

decreased. For example, in PATCH surfaces, the average heritability between IB 4 and

IB 32 was fixed at 0.25 for low correlations, in contrast with a change from 0.29 to 0.35

for the high correlations set (Figure 3-3). Thus, the use of smaller incomplete blocks

should be preferred for field experiments under the presence of moderate to strong spatial

correlations.


0.40 -

0.38 -

0.36 -

0.34 -

0.32 -

0.30 -

0.28 -

0.26 -

0.24


CR RCB IB4 IB8 IB16 IB32 R-C


Figure 3-3. Average individual heritability estimates for ALL and PATCH surface pattern
for simulated surfaces for subsets of simulations with High (px > 0.7 and py >
0.7) and Low (px < 0.3 and py < 0.3) spatial correlations.









Latinization

A z-test was conducted to compare average H2 values for all designs and surface

patterns generated from simulations with and without Latinization but always fitted

without the Latinization random effects (i.e., Partial-LAT versus No-LAT). This test

indicated significant differences (experiment-wise a= 0.05) for all IB and R-C designs

on PATCH surfaces only. In this surface pattern, Latinized designs produced an increase

up to 0.04 units of heritability, which tended to be larger for simpler designs. It was

expected that the incorporation of a restricted randomization, as occurs with Latinized

designs, would yield smaller standard deviation for HB: and for the variance component

o2block(rep) (Williams et al. 1999); but when Partial-LAT was compared to No-LAT very

small reductions in heritability standard deviations were found and changes in coefficient

of variation were not relevant (for H2: see Table A-7).

The other comparison considered only simulated datasets originating from

Latinized experiments, which were fitted using linear models with and without

Latinization effects (i.e., Full-LAT and Partial-LAT). The results from the likelihood

ratio test indicated statistical significant differences (experiment-wise a= 0.05) between

these two models for all designs and surface patterns with the exception of designs IB 4

and IB 8 for PATCH surfaces. Better model fitting was found on Full-LAT, but

differences in terms of average heritability, clonal and error variance were very small and

not of practical relevance. Also, the largest improvements of including Latinization

effects occurred in GRAD surfaces or with IB 32 designs in any surface. The above

findings agree with what Qiao et al. (2000) reported for the analysis of several Latinized

R-C designs, where greater efficiencies were reported in those analyses that included the









Latinized effects in the linear model. In summary, once a design is generated with

Latinization, the incorporation of its corresponding random effects is important.

The small benefits of Latinization found in this study in the design and analysis

phase were not important in practice, but could possible be a consequence of the large

sample size used per clone (8 ramets). Nevertheless, the use of Latinization in the design

stage is recommended to help control potential undetected environmental biases.

Number of Ramets per Clone

The variance components estimated through REML for all experimental designs

and surface patterns were very similar for datasets with different number of ramets per

clone. Hence, even with reduced amounts of information, REML yielded unbiased

variance component estimates and best linear unbiased predictions of genetic values. The

main impact of larger experiments (i.e., more ramets per clone) was a decrease in

variability or improved precision on variance component estimates. The greatest overall

reduction occurred when the number of ramets changed from 2 to 4, and was more

pronounced for PATCH surfaces. For example, within R-C designs the coefficient of

variation for the clonal variance component on average in all surfaces was reduced from

22.5% for analyses with 2 ramets to 12.0% for the case with 8 ramets (Table 3-4). As

with variance components, for the same design type and surface pattern average

heritability estimates changed very little between datasets with different numbers of

ramets. Nevertheless, HB distributions for any design type and surface pattern showed a

decrease in range (and variance) as the number of ramets increased (for R-C design see

Figure 3-4).











Table 3-4. Coefficient of variation (100 std / x) for the estimated variance component for
clone and error on 3 surface patterns, different number of ramets per clone and
3 selected experimental designs.

Clonal Variance
ramets / clone

Design a Surface b 2 4 6 8

RCB ALL 21.5 % 14.4 % 13.1% 12.0 %

GRAD 23.6 % 14.6 % 12.7 % 11.7 %

PATCH 24.5 % 15.4 % 12.7 % 11.8 %

IB 32 ALL 21.5 % 14.5 % 12.8 % 11.7 %

GRAD 21.3 % 14.4 % 11.9 % 11.3 %

PATCH 25.3 % 15.7 % 13.6 % 12.7 %

R-C ALL 21.5 % 15.3 % 13.0 % 12.1%

GRAD 21.2 % 14.3 % 12.6 % 11.9 %

PATCH 24.8 % 15.0 % 13.2 % 12.0 %

Error Variance
ramets / clone

Design Surface 2 4 6 8

RCB ALL 15.0% 12.2% 11.0% 9.8%

GRAD 15.9% 11.1% 9.9% 8.9%

PATCH 10.5% 7.1% 5.0% 4.6%

IB 32 ALL 22.0% 20.3% 20.0% 19.6%

GRAD 14.9% 12.6% 12.3% 12.2%

PATCH 17.1% 15.3% 15.0% 14.7%

R-C ALL 22.5% 21.3% 20.8% 20.5%

GRAD 14.1% 11.9% 11.6% 11.6%

PATCH 18.9% 17.8% 17.3% 17.2%
a RCB, randomized complete block; IB 32, incomplete block with 32 blocks per replicate; R-C, row-
column.
b ALL, surfaces with patches and gradients; GRAD, surfaces with only gradients; PATCH, surfaces with
only patches.














Surface: ALL


160

140

120

> 100

3 80

LL 60

40

20

0


Surface: PATCH


160

140

120

> 100-
a,
3 80

1 60

40

20 -

0
/

0.0 0.1


Figure 3-4. Distribution of individual broad-sense heritability estimates (H2 ) for row-

column designs with 2 and 8 ramets per clone obtained from 500 simulations
each for the surface patterns ALL and PATCH from fitting experimental
designs and analyses that were non-Latinized (No-LAT).


020


018
\
% 016
o
> 014
5
012
I

w 010


008 -
C-

0 06


004
8 2-4


ramets / clone


4-6
ramets / clone


Figure 3-5. Average values for row-columns design in all surface patterns from fitting
experimental designs and analyses that were non-Latinized: a) Clonal mean
heritabilities (the whiskers correspond to the 95% confidence intervals); and
b) Clonal mean heritability increments.


0.2 0.3 0.4 0.5 0.6 0.7
HB2


0.3 C
HB2









For all studied designs, as the number of ramets increased considerably higher

clonal mean heritabilities (H/2) were found; nevertheless, incremental improvements

above 6 ramets per clone were small. For R-C designs Hf2 averaged over all surfaces for

2, 4, 6 and 8 ramets per clone were 0.44, 0.62, 0.70 and 0.76, respectively (Figure 3-5a);

and the largest change occurred when the number of ramets changed from 2 to 4.

Notably, these increments were always larger in PATCH surfaces than any other surface

(see Figure 3-5b).

Hence, the results from these simulations (that considered no missing values)

indicated that the optimal number of ramets per clone per site should be between 4 and 6.

This number of ramets agrees with results obtained in other studies that recommend

between 1 and 6 ramets per clone per site (Shaw and Hood 1985; Russell and Libby

1986), and will produce near maximal genetic gains for clonal selection and will allow

alternative uses of available resources (for example, increasing the number of clones

tested).

Conclusions and Recommendations

Appropriate selection of experimental designs can yield considerably

improvements in clonal testing. Row-column and incomplete block designs with a block

size of 8 trees had the higher individual heritability estimates. Unfortunately, these two

designs yielded more variable heritability estimates than simpler designs.

Approximate 95% confidence intervals on heritability estimates obtained using

Dickerson's method were similar to empirical percentile CI's. These methods were

particularly close for simple designs (CR and RCB), but Dickerson's approximation

presented a bias in the upper confidence limit of more complex designs (IB and R-C).









The examination of the simulation parameters used to generate the surfaces patterns

had larger individual heritabilities when the amount of white noise or unstructured

residual 2ms was smaller and the spatial correlations larger, with a greater effect of the

latter at low levels of C2ms. Also, the effect of different levels of gradients on Hf2 tended

to be almost non-existent, affecting slightly GRAD surfaces only.

In the present study, the use of Latinized designs improved the experimental design

efficiency mainly for PATCH surfaces, but the benefits of designing an experiment with

Latinization were not important in practical terms. On the other hand, once an experiment

is designed with Latinization, the inclusion of its corresponding effects in the linear

model yielded better analyses.

As expected, experiments with more ramets per clone produced more precise

variance component estimates and larger clonal mean heritabilities. Using 4 to 6 ramets

per clone per site is recommended. More than 6 ramets produced marginal improvements

in precision of clonal means, but for clonal forestry it is important to screen large

numbers of clones in one test.

Finally, it is important to note, that these findings were produced analyzing datasets

that did not have missing values. Nevertheless, in other related studies (see Chapter 2 and

Fu et al. 1999b) the effects of mortality were small. Therefore, these results can be

extended safely to situations with low mortality levels.















CHAPTER 4
ACCOUNTING FOR SPATIAL VARIABILITY IN BREEDING TRIALS

Introduction

The amount and type of environmental heterogeneity found in agricultural or

forestry trials greatly influences the statistical precision obtained in the comparison of

treatments. As the homogeneity decreases, the error variance of the treatment effect

estimates increases and, consequently, it is harder to detect differences among treatments.

Many natural and anthropogenic factors affecting site variability, such as soil,

topography, wind, machinery or planting technique are usually detected after test

establishment, making it difficult to implement optimal designs and to minimize the

portion of the experimental error due to site variation (Fu et al. 1999). This

environmental or spatial variability usually is expressed as gradients, patches or a

combination of both, together with random microsite differences (Costa e Silva et al.

2001).

Proper treatment randomization is enough to ensure that, over repeated testing and

sampling, unbiased estimates of treatment effects are obtained, and hypothesis testing is

valid (Grondona and Cressie 1991; Brownie et al. 1993). Nevertheless, under

environmental heterogeneity, the estimates obtained for the analysis of a particular

dataset are highly variable. This situation can be improved by the use of proper

experimental designs and statistical analysis. Randomized complete block, incomplete

block and row-column designs constitute classical approaches that control for spatial

variability and are recognized as "a priori" techniques because they are implemented in









the design stage. Another group of techniques ("a posteriori ") deals with statistical

analyses that incorporate the x and y coordinates of the experimental units (plots, trees or

plants) in the linear model to account for physical proximity.

The majority of statistical analyses for spatial data are based on modeling two main

components of spatial variability (Grondona et al. 1996): 1) global trends or large-scale

variation; and 2) local trends or small-scale variation. The first component corresponds to

field gradients which can be modeled by: i) incorporating designs effects (e.g., replicate,

block, row, column, etc.); ii) using x and y positions as one or multiple fixed effects in

the linear model to describe continuous mathematical functions; or, iii) calculating

covariates to adjust each observation. Some of the most common options are 1)

polynomial regression (Brownie et al. 1993); 2) cubic smoothing spline (Durban et al.

1999; Verbyla et al. 1999); 3) nearest neighbor analysis (Bartlett 1978; Brownie et al.

1993; Vollmann et al. 1996); and 4) moving average (Townley-Smith and Hurd 1973).

Of particular interest are the nearest neighbor techniques, for which there are multiple

variants. Nevertheless, the majority corresponds to modifications of the Papadakis

method (Atkinson 1969) which is based on the use of one or more covariates obtained

from averaging residuals of neighboring plots; these residuals are obtained from fitting a

completely randomized or randomized complete block design. The original method uses

the residual average of one plot to the right and one to the left; thus, known as EW

adjustment. Several modifications are available which consider different definitions of

the covariates with 2, 4, 8 or even 24 adjacent plots (Brownie et al. 1993; Stroup et al.

1994; Vollman et al. 1996). Also, Bartlett (1978) suggested performing an iterated

Papadakis method that recalculates residuals iteratively. A method similar to Papadakis is









based on the original measurements and not on residuals is the Moving Average analysis.

Here, a "local" mean is calculated from a number of adjacent plots and used as a

covariate (Townley-Smith and Hurd, 1973).

The second component of spatial variability (local trends), frequently identified as

multiple patches on the field surface, can be modeled through the specification of an error

structure that takes into account some form of spatial correlation produced by the

physical proximity among experimental units. Most of the techniques for modeling global

trends (as presented above) assume that residuals or errors from the linear model are

independent and identically distributed with a common variance. Under spatial

correlation, the error variance-covariance matrix has nonzero off-diagonal elements,

which are assumed to be a function of the distance among experimental units.

Several correlated error models are available (Cressie 1993; Littell et al. 1996),

allowing substantial flexibility in assumptions and covariance structures; nevertheless,

different error structures tend to yield similar results (Zimmerman and Harville 1991).

The first-order separable autoregressive error structure has been used successfully in

several agronomic and forestry trials (Grondona et al. 1996; Gilmour et al. 1997; Qiao et

al. 2000; Sarker et al. 2001; Costa e Silva et al. 2001; Dutkowski et al. 2002). This

pattern considers two perpendicular correlations (one for row and the other for the

column direction), and is equivalent to the separable exponential geostatistical model.

Sometimes a nugget parameter is included in the error structure to model the expected

variability that occurs if repeated measurements were made exactly at the same location

(Cressie, 1993) allowing modeling of potential microsite variability or measurement

error.









Finally, it is possible to fit global and local trends simultaneously for which

multiple combinations of techniques exist. Because many of these options yield similar

results, it is more difficult to select an adequate model, and the use of some statistical and

graphical tools (e.g., sample variograms) seems particularly useful (Gilmour et al. 1997).

Field testing of varieties is critical for the progress of genetic improvement

programs, and is an expensive and time consuming activity. Therefore, allocation of

resources and the use of the best available statistical techniques to obtain precise

estimation of genetic parameters, such as heritability and breeding values, are critical.

Due to the large number of varieties or treatments to be studied relatively large test areas

are required. These areas usually have high amounts of environmental heterogeneity that

inflate the residual error variance. Hence, controlling for this variability is important, and

requires the use of more sophisticated experimental designs, or the implementation of

spatial statistical analyses.

The goal of this study is to quantify efficiencies of implementing a range of "a

priori" and "a posteriori" statistical techniques to account for spatial variability, and also

to discriminate among these techniques for parsimony and global performance for a

broad set of conditions. In particular, the selected array of techniques is compared to

improvement in the precision for comparing and predicting treatment effects. This is

achieved through simulations of field trials with sets of unrelated genotypes (or

treatments) tested on a single site with environments having different patterns of

heterogeneity (only patchy, only gradient, and both types). These patterns were generated

using a polynomial function for the gradients and an AR1OAR1 error structure for

patches. The selected techniques included 1) modeling of global trends through the use of









traditional experimental designs (complete randomized, randomized complete block,

incomplete block and row-column) or polynomial models; 2) fitting AR1AR1 error

structure (with and without nugget) individually or in combination with some of the

above specifications of global trends; and 3) implementation of several variants of

Papadakis and Moving Average methods.

Materials and Methods

Field Design and Simulation

The simulations were based on a single trial of 2,048 plots arranged on a

contiguous rectangular site of 64 rows and 32 columns with square spacing and no

missing observations. For each site a total of 256 treatments genotypess or varieties) with

8 replicates each were tested in plots composed by single plants. The linear model used to

generate the simulated data for the response variable was

Y ijk = P+ gk+ Es(jk) + Ems(lJk) 4-1

where yijk is the response of the plant located in the ith row and jth column of the kth

treatment, u is a fixed population mean which was set equal to 10 units, gk is the random

treatment effect, Es(ijk) is the surface error (or structured residual) and Ems(ijk) corresponds

to the microsite random error (or unstructured residual sometimes called nugget). The

total variance for the above linear model is C2T = C2g + C2s + c2ms. For simplicity, but

without loss of generality, y2T was fixed to 1, and the variances for all surfaces were set

to 02, = 0.25 and 2s + c2ms = 0.75.

The base experimental design simulated corresponded to an incomplete block

design with 32 blocks in each of the 8 resolvable replications in an a-lattice design (John

and Williams 1995, p. 68), which had previously shown to be superior to completely









randomized and randomized complete block designs (Chapter 2). This experimental

design was implemented in three surface patterns: only patches (PATCH), only gradients

(GRAD), and both types (ALL). These simulated error surfaces included two main

components: gradients and patches. The following third degree polynomial function

using the x and y coordinates of each plot was used to generate the gradients:

t' = a (x +Y,) +y+3(xc 2y + xcly ) 4-2

where xc and yci correspond to the centered position values for the ith plot (i.e., xci = xi -

x) located in column x and row y. Patches were generated using a first-order separable

autoregressive error structure (AR1OAR1) based on two perpendicular spatial

correlations (px and py) with an error variance-covariance matrix defined as

Var (e) = cs + cms for diagonal elements 4-3

Cov (e,, e,) = os pxhxp hy for off-diagonal elements 4-4

where hx = I xi xi' 1, hy = I yi yi' |, i.e., the absolute distance in the row and column

position, respectively, between two plots.

The simulation consisted of the generation of independent surfaces over which

different experimental designs were superimposed. Each surface was produced by

selecting 5 parameters at random from uniform distributions (a, /f, px, py and c2ms). The

parameters were restricted to the following ranges: 0 to 0.05 for a and 0 to 0.0005 for /Y,

0.01 to 0.99 for the correlation parameters px and py, and 0.15 to 0.60 for 2 Ms.

Additionally, the two correlations were restricted so their absolute difference was smaller

than 0.85 and C2y was calculated from C2y = 0.75 G2Ms. Several layouts of the

incomplete block design were generated using the software CycDesigN and randomly

selected (Whitaker et al. 2002). For each surface pattern 1,000 simulations were









generated and stored for further analysis. More details of the simulation process can be

found in Chapter 2.

Statistical Analysis

For the prediction of treatment effects and estimation of other parameters of interest

four sets of statistical analyses were considered based on modeling: 1) only global trends;

2) only local trends; 3) both components simultaneously; and 4) nearest neighbor

methods. All linear models were fitted using the software ASREML (Gilmour et al.

2002), which estimates fixed effects, variance-covariance parameters and predicts

random effects by REML (Patterson and Thompson 1971).

For those analyses that considered only global trends it was assumed that the errors

were independent and identically distributed; therefore, no spatial error structure was

considered. To model the global trends or mean structure, classical experimental designs

and explicit trend modeling were implemented. The classical designs or models

corresponded to a completely randomized (CR), randomized complete block (RCB), the

original incomplete block design with 32 blocks (IB) which was employed to simulate

the data, and a superimposed row-column (R-C) design that consisted of short rows and

columns within each replicate. All effects with the exception of the overall mean were

considered random (Table 4-1). The original a-lattice design did not consider row and

column effects; nevertheless, it is of interest to include this type of analysis because the

short rows and columns allow for some form of spatial modeling (Lin et al. 1993).

Explicit trend modeling considered a set of continuous variables to describe polynomial

regressions. The independent variables used in the study corresponded to the x and y

coordinates of the individual plots, which were assumed fixed. The models used are










specified on Table 4-2. It is important to note that the model Full-Poly was the one used

to generate the gradients in the error surface simulations (see Equation 4-2).

For modeling the local trends two variants of the AR1AR1 error structure were

fitted with and without nugget effect. The variant with nugget is specified by Equations

4-3 and 4-4; and for the case of no nugget, the only difference is that the variance

component C2 ms was assumed to be zero. These two variants were fitted alone or in

combination with different forms of global trends: 1) the classical models specified in

Table 4-1; and 2) the polynomial regressions from Table 4-2.



Table 4-1. Linear models fitted for simulated datasets using classical experimental
designs to model global trends: complete randomized (CR), randomized
complete block (RCB), incomplete block design with 32 blocks (IB), and row-
column (R-C) design. All model effects other than the mean were considered
random.

Model a
CR Yj = P + g, + +1
RCB yj = p + rep, + gj + ~1
IB yjk = + rep + block(rep), + gk + Fk
R-C Yjkl = + rep1 + col(rep), + row(rep)lk + gl + Flkl
a g, treatment effect; rep, resolvable replicates; block(rep), incomplete block nested within replicate;
col(rep), column nested within replicate (short column); row(rep), row nested within replicate (short
row); and E, residual.


Table 4-2. Linear models fitted for simulated datasets using polynomial functions to
model global trends: linear (Linear), quadratic (Red-Poly), and quadratic
model with some interactions (Full-Poly). All variables are assumed fixed and
the treatment effect is random.

Model a
Linear Yljk = + PO X, + p1 Yj + gk + k jk
Red-Poly yljk = + po Xl + p1 YJ + f2 X2 + 3 Yj2 + gk + Eljk
Full-Poly Yljk = X + Po X1 + P1 Yj + P2 X2 Yj + P3 Yj2 X1 + gk + jk
a x, longitudinal position of the plant; y, latitudinal position; g, treatment; and e, residual.









For the nearest neighbor methods, a number of different variants of Papadakis

(PAP) and Moving Average (MA) methods were implemented. The design effects

(replicate, block, etc.) were not considered when these methods were implemented. The

residuals (or deviations) used to implement the PAP method and its variants were

computed as ,y = y, wk, where y, is the observation of the plot located in the ith row and

jth column and yk is the simple treatment average (fitting a CR design). The covariates

used to correct for differential yields (X,j) were calculated using the average of these

residuals from a number of neighboring plots. In the case of the MA method, the

covariate was calculated using the original observations (y,) instead of their residuals

(n, ). The following linear model, which considered the covariates as fixed effects and the

treatment effects (gk) as random, was used

y,jk -= + Po X1,1, + P X2,j + ... + m Xp,,j + gk+ ,jk 4-5

where Xi,,, X2,, ..., Xp,, are p different covariates. All definitions of covariates and linear

models considered for this study are detailed in Figure 4-1. All Papadakis or Moving

Average models were fitted with an error structure assuming independent errors. Also, an

iterated PAP method was implemented for the models PAP-6, PAP-8 and PAP-11. Three

iterations were implemented on these models using the residuals obtained from the

previous iteration.

In summary, a total of 35 different statistical models where fitted for 3 difference

surface patterns (PATCH, GRAD and ALL) each one with 1,000 simulated datasets

totaling 105,000 sets of estimated parameters; additionally 24,000 sets were available

from the iterative PAP models.








58



The output from each of the fitted models, datasets and surface patterns was


compiled and summarized to obtain averages for variance component parameters. Also,


empirical correlations between true and predicted treatment effect values (CORR) were


calculated to compare different techniques and models.


PAP-1








PAP-4


PAP-4


PAP-7 / MA-1


1


1


PAP-2
2

1




2


PAP-5


PAP-8 / MA-2


1 1 1


1 1 1
1 1 l


PAP-10


PAP-3




1 1


PAP-11

2

5 1 6

3 3

6 1 5

2


PAP-6

2

1

3 3

2

2


PAP-9 / MA-3


212
2 1 2

1 1

2 1 2


Figure 4-1. Neighbor plots and definitions of covariates used in Papadakis (PAP) and
Moving Average (MA) methods. Plots with the same numbers indicate a
common covariate.









Results

Modeling Global and Local Trends

Comparisons of the average correlation values between true and predicted

treatment effects showed considerable differences among linear models, error structures,

and surface patterns with values that ranged from 0.85 to 0.90 (Figure 4-2). While this

range is relatively narrow, small increments in values closer to 1 correspond to large

improvements in terms of precision. In relation to those models that assumed independent

errors (ID), all designs yielded higher average CORR values than CR. In general, the best

model was R-C, but Full-Poly was better for surfaces with only global gradients

(GRAD). In general, the models with explicit trends (Linear, Red-Poly and Full-Poly)

were very good for GRAD surfaces and unsuccessful for sites with only local patches

(PATCH). Nevertheless, for patchy surfaces polynomial models were slightly better than

CR, which is an indication that a small portion of the patches can be explained as trends.

Approximately 98% and 95% of the fitted models converged for the autoregressive

without nugget (AR1OAR1) and with nugget (AR1AR1+r) error structures,

respectively. In GRAD surfaces, the results for the latter structure are not reported

because very few simulations converged. For all surface patterns, most of the failed

convergences occurred in surfaces without patches, i.e., when the spatial correlations (px

and py) were close to zero. This situation makes the covariance formula in Equation 4-4

equal to zero; hence, the variance of the error composed by Cy2 and Y2ms as indicated in

Equation 4-3 cannot be adequately partitioned.

Once the errors were assumed correlated (i.e., when the error structure was

incorporated in the linear models) several changes occurred (Figure 4-2). First,









AR1OAR1 had for all models on surfaces ALL and PATCH larger average CORR values

than model with independent error structure. The largest improvement in CORR was

found on PATCH surfaces, which was expected because in this case now the error

structure is explicitly modeled. Also, the CORR values in GRAD surfaces almost did not

change between AR1IAR1 and ID error structures, with the exception of CR designs

that had better CORR values than with ID.

Once the nugget effect was incorporated, as with AR1AR1+rj, even larger CORR

values were found. Comparisons of this error structure with the other two showed that

differences between experimental designs were even smaller. The maximum values for

CORR were 0.90 and 0.88 with the Full-Poly model for ALL and PATCH, respectively.

But it is important to note that this last model corresponds exactly to the one used to

simulate the data; therefore, in this study it was used as reference or base model for the

maximum performance that can be achieved.

As with CORR, the average estimated variance components differed between error

structures and models (see Tables 4-3, 4-4 and 4-5). And, as expected, within the same

error structure, there was an inverse relationship between the variance of the error and the

average CORR value. For all designs, surfaces and error structures, the mean treatment

variance component (C2g) after 1,000 simulations was close to its parametric value of

0.25; and, hence, unbiased. In addition, the estimate of the error variance ( 2 or c2 +

C2ms) was near its parametric value f 0.75 for all surface patterns and error structures (for

models that partitions the error structure into replicates, blocks, rows and columns, these

components are included in the total sum of 0.75).










ID AR10 AR1 AR1 AR1+Tl
0.90


0.89


0.88





0.86 -- PATCH


0.85




Figure 4-2. Average correlations between true and predicted treatment effects (CORR) in
3 different surface patterns for classical experimental design analyses and
polynomial models fitted for the following error structures: independent errors
(ID); autoregressive without nugget (AR1OAR1); and autoregressive with
nugget (AR1AR1+rI). Surface GRAD is not shown in the last graph because
few simulations converged.




The average variance components for the reference model from this simulation

study (Full-Poly with AR1AR1+rq) for PATCH surfaces were close to their parametric

constants, with the exception of px which was slightly overestimated (0.52 instead of

0.50). A similar situation was found in ALL, but in this case the error variance only sums

up to 0.56 (instead of the parametric 0.75) because part of the variability was explained

by the polynomial fixed effects. Hence, for this surface explicit trends explained

approximately 19% of the total variability.

Within AR1AR1+qr important differences were found when a model different

than Full-Poly was used. In ALL surfaces, larger average px and py were obtained with

values as high as 0.88. Because of the lack of trends in PATCH the spatial correlation for







62


other models different than Full-Poly were close to 0.50. In the case of GRAD, the

correlations for AR1IAR1 should be close to zero, but in almost all models, slightly

positive values were found with the exception of Full-Poly. This bias was particularly

large for CR, possible a consequence of failing to model the underlying surface trends

properly.



Table 4-3. Average variance components for the surface pattern GRAD obtained from
fitting the models in Tables 4-1 and 4-2 for the following error structures: a)
independent errors (ID); and b) autoregressive without nugget (AR1 AR1).
All values are the means from 1,000 simulations and the true values were C2
=0.25 and G2 + C2ms = 0.75 for CR design.

ID
Model g b rep block(rep) col(rep) row(rep) px Py 72ms (2s
CR 0.25 0.75
RCB 0.25 0.16 0.61
IB 32 0.25 0.16 0.05 0.56
R-C 0.25 0.16 0.04 0.02 0.55
Linear 0.25 0.60
Red-Poly 0.25 0.60
Full-Poly 0.25 0.56

ARIAR1
Model g rep block(rep) col(rep) row(rep) px py 72ms (2s
CR 0.25 0.19 0.20 0.74
RCB 0.25 0.16 0.07 0.07 0.61
IB 32 0.25 0.16 0.04 0.03 0.02 0.57
R-C 0.25 0.16 0.04 0.02 0.02 0.01 0.56
Linear 0.25 0.07 0.07 0.60
Red-Poly 0.25 0.06 0.07 0.60
Full-Poly 0.25 0.00 0.00 0.56
a CR, complete randomized; RCB, randomized complete block; IB 32 incomplete block with 32 blocks per
replicate; R-C, row-column; Linear, linear regression; Red-Poly, reduced polynomial regression; Full-
Poly, full polynomial regression.
b g, treatment; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep),
column nested within replicate; row(rep), row nested within replicate; px, correlation in the x-direction;
py, correlation in the y-direction; ms, microsite error; s, surface error.












Table 4-4. Average variance components for the surface pattern ALL obtained from
fitting the models in Tables 4-1 and 4-2 for the following error structures: a)
independent errors (ID); b) autoregressive without nugget (AR1OAR1); and
c) autoregressive with nugget (AR1OAR1+rj). All values are the means from
1,000 simulations and the true values were C2 = 0.25 and 2s + C2ms = 0.75
for CR design.

ID
Model g b rep block(rep) col(rep) row(rep) px Py (2ms 2s
CR 0.25 0.75
RCB 0.25 0.17 0.60
IB 32 0.25 0.17 0.09 0.51
R-C 0.25 0.17 0.06 0.05 0.50
Linear 0.25 0.61
Red-Poly 0.25 0.60
Full-Poly 0.25 0.55

ARIAR1
Model g rep block(rep) col(rep) row(rep) px Py (2ms 2s
CR 0.25 0.29 0.28 0.72
RCB 0.25 0.16 0.19 0.18 0.60
IB 32 0.25 0.16 0.02 0.16 0.13 0.58
R-C 0.25 0.16 0.04 0.03 0.12 0.12 0.53
Linear 0.25 0.20 0.19 0.60
Red-Poly 0.25 0.19 0.18 0.60
Full-Poly 0.25 0.15 0.14 0.55

ARIOARI+rT
Model g rep block(rep) col(rep) row(rep) px Py (2ms 2s
CR 0.25 0.87 0.88 0.44 0.33
RCB 0.25 0.07 0.76 0.78 0.35 0.35
IB 32 0.25 0.06 0.00 0.79 0.79 0.42 0.29
R-C 0.25 0.07 0.00 0.00 0.75 0.77 0.35 0.33
Linear 0.25 0.73 0.72 0.41 0.22
Red-Poly 0.25 0.73 0.72 0.40 0.22
Full-Poly 0.25 0.54 0.51 0.36 0.20
a CR, complete randomized; RCB, randomized complete block; IB 32 incomplete block with 32 blocks per
replicate; R-C, row-column; Linear, linear regression; Red-Poly, reduced polynomial regression; Full-
Poly, full polynomial regression.
b g, treatment; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep),
column nested within replicate; row(rep), row nested within replicate; px, correlation in the x-direction;
py, correlation in the y-direction; ms, microsite error; s, surface error.











Table 4-5. Average variance components for the surface pattern PATCH obtained from
fitting the models in Tables 4-1 and 4-2 for the following error structures: a)
independent errors (ID); b) autoregressive without nugget (AR1OAR1); and
c) autoregressive with nugget (AR1OAR1+rj). All values are the means from
1,000 simulations and the true values were C2 = 0.25 and 2s + C2ms = 0.75
for CR design.

ID
Model g b rep block(rep) col(rep) row(rep) px Py (2ms 2s
CR 0.25 0.75
RCB 0.25 0.02 0.73
IB 32 0.25 0.02 0.10 0.63
R-C 0.25 0.02 0.06 0.06 0.62
Linear 0.25 0.74
Red-Poly 0.25 0.74
Full-Poly 0.25 0.73

ARIAR1
Model g rep block(rep) col(rep) row(rep) px py (2 (72s
CR 0.25 0.22 0.21 0.74
RCB 0.25 0.02 0.21 0.20 0.73
IB 32 0.25 0.01 0.02 0.20 0.18 0.71
R-C 0.25 0.02 0.04 0.03 0.17 0.16 0.66
Linear 0.25 0.22 0.21 0.73
Red-Poly 0.25 0.22 0.21 0.73
Full-Poly 0.25 0.21 0.20 0.73

ARIOARI+rT
Model g rep block(rep) col(rep) row(rep) px py (2ms 2s
CR 0.25 0.52 0.50 0.37 0.39
RCB 0.25 0.00 0.53 0.51 0.33 0.43
IB 32 0.25 0.00 0.00 0.51 0.49 0.36 0.39
R-C 0.25 0.00 0.00 0.00 0.51 0.50 0.33 0.42
Linear 0.25 0.52 0.50 0.37 0.40
Red-Poly 0.25 0.52 0.50 0.36 0.40
Full-Poly 0.25 0.52 0.50 0.37 0.40
a CR, complete randomized; RCB, randomized complete block; IB 32 incomplete block with 32 blocks per
replicate; R-C, row-column; Linear, linear regression; Red-Poly, reduced polynomial regression; Full-
Poly, full polynomial regression.
b g, treatment; rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep),
column nested within replicate; row(rep), row nested within replicate; px, correlation in the x-direction;
py, correlation in the y-direction; ms, microsite error; s, surface error.









Considering only models with independent errors, the partition of variance

components showed larger values of the replicate variance component in ALL and

GRAD in comparison with PATCH. Also, the incomplete block variance components

(from IB and R-C) were always smaller in GRAD surfaces.

Because few simulations converged for GRAD surfaces, to draw inferences

comparisons between AR1IAR1 and AR1AR1+rj can only be performed for surfaces

ALL and PATCH. Failing to incorporate the nugget effect in these surface patterns

produced an under estimation of the spatial correlation parameters for any experimental

design. Another consequence was an increase in the magnitude of the variance

components for other effects. For example, in ALL surfaces the replicate variance

increased from 0.07 to 0.16, indicating that part of the variability that the error structured

failed to capture was absorbed by the replicate effect.

Nearest Neighbor Models

Due to differences in the number of neighbor plots accounted for and covariates

considered, a wide variety of average CORR values were found for the nearest neighbor

methods (Figure 4-3), but in all cases, average CORR values greater than 0.86 were

found. In general, models that considered more plots and/or covariates produced larger

CORR values. When compared with analyses with RCB designs, both the PAP and MA

methods gave larger CORR values for PATCH and ALL, but for GRAD surface RCB

was in general better than any nearest neighbor methods. This situation is not surprising

because PAP and MA were originally designed to control for correlated errors (or

patches) and not potential surface trends or gradients. Nevertheless, as occurred before

with the incorporation of autoregressive error structures, controlling for patches helped to









model a small portion of the trends (see GRAD in Figures 4-2 and 4-3). Additionally, for

all nearest neighbor methods studied, unbiased estimates of variance components were

always obtained, as indicated by values close to 0.25 for all cases. Also, a proportional

inverse relation was noted between the average error variance and the CORR values.

Generally, PAP-6 and PAP-11 were superior to any of the MA methods across all surface

patterns (Figure 4-3), where PAP-11 was slightly better, particularly for GRAD surfaces.

Interestingly, the PAP-8 method that only considers one covariate for the 8 surrounding

plots was particularly good for GRAD.

PAP-6, PAP-8 and PAP-11 were also fitted iteratively using their previously

estimated residuals to calculate new covariates. For any given surface, a decrease in the

average CORR and the treatment and error variance was noted in PAP-6 and PAP-11,

this decrease was more abrupt in the last iteration (Table 4-6) and affected mostly

surfaces with patches (ALL and PATCH). PAP-8 was almost not affected by the

iterations showing stable statistics and unbiased variance components.

A group of selected models was plotted in Figure 4-4. The first 4 models assumed

that the errors were independent (ID) and the last 2 models corresponded to the best

models for the autoregressive error structures. ID models did not achieve the maximum

potential CORR values that could be obtained by using the Full-Poly with AR1ARl+rq.

As indicated before, the classical design R-C was very good in ALL and GRAD, but less

precise for PATCH surfaces than other models. PAP-6 and PAP-11 performed very well

but had slightly lower CORR values than the maximum potential CORR obtained with

the ARlOARl+r.



















0.88 -


0.87 -


0.86- -0- ALL
GRAD
-- PATCH

0.85 i, ,



Figure 4-3. Average correlations between true and predicted treatment effects (CORR) in
3 different surface patterns for nearest neighbor analyses fitted assuming
independent errors. PAP: Papadakis, MA: Moving Average.


RCB R-C PAP-6 PAP-11 Full-Poly Full-Poly
(ID) (ID) (ID) (ID) (AR1gAR1) (AR1gAR1+q1)


Figure 4-4. Average correlations between true and predicted treatment effects (CORR) in
3 different surface patterns for selected methods: randomized complete block
(RCB), row-column (R-C), and Papadakis (PAP) using the following error
structures: independent errors (ID), and autoregressive with nugget
(ARIAR1+r).










Table 4-6. Average correlation between true and predicted treatment effects (CORR) and
average treatment variance component for iterated Papadakis methods PAP-6,
PAP-8 and PAP-11 for the surface patterns ALL, PATCH and GRAD.

ALL a
PAP-6 PAP-8 PAP-11
Iteration CORR 0gb 2, CORR o a2 CORR o 2,
0 0.89 0.25 0.52 0.89 0.25 0.54 0.89 0.25 0.51
1 0.89 0.25 0.52 0.89 0.25 0.53 0.89 0.26 0.50
2 0.89 0.25 0.47 0.89 0.25 0.53 0.89 0.25 0.44
3 0.87 0.24 0.45 0.89 0.25 0.53 0.87 0.24 0.39

PATCH
PAP-6 PAP-8 PAP-11
Iteration CORR og, 2, CORR o, 0G2 CORR o 2,
0 0.88 0.25 0.59 0.87 0.25 0.65 0.88 0.25 0.59
1 0.86 0.26 0.62 0.87 0.25 0.64 0.87 0.26 0.57
2 0.87 0.25 0.49 0.87 0.25 0.64 0.87 0.25 0.47
3 0.83 0.23 0.49 0.87 0.25 0.64 0.85 0.24 0.44

GRAD
PAP-6 PAP-8 PAP-11

Iteration CORR 02, a2 CORR 02, a2 CORR o2 02
0 0.88 0.25 0.61 0.88 0.25 0.60 0.88 0.25 0.60
1 0.88 0.25 0.60 0.88 0.25 0.60 0.88 0.25 0.59
2 0.87 0.25 0.59 0.88 0.25 0.60 0.87 0.25 0.56
3 0.87 0.24 0.56 0.88 0.25 0.60 0.86 0.25 0.45
a ALL, surfaces with patches and gradients; GRAD, surfaces with only gradients; PATCH, surfaces with
only patches.
b g, treatment effect; s, surface error.


Discussion

Sites for field trials in agronomy or forestry are particularly variable showing

spatial heterogeneity that produces positive correlations between plots in close proximity.

In this context, classical experimental designs that assume independent errors are still

valid due to the randomization of treatments, but they are not optimal. Some of the

consequences are (Ball et al. 1993) 1) inflation of the mean square error (MSE); 2) MSE











is not the minimum-variance estimator of the error; 3) overestimated standard error of

treatment pairwise comparisons; 4) invalid hypothesis testing and confidence intervals;

and 5) bias in other statistics calculated from variance components estimates (e.g.,

heritability and genetic correlations). Nevertheless, controlling for spatial variation by

modeling the error structure or using nearest neighbor analysis helps to minimize these

problems and to obtain more efficient analyses.

A concern in this study was CORR values close to the maximum of 1, which was

probably a consequence of using a large number of replicates in the field design. To study

this, the first 500 simulated dataset from the ALL surface were fitted using only 2

contiguous replicates and the models from Tables 4-1 and 4-2 with all three error

structures described previously. The results showed an average CORR for all surfaces

that ranged from 0.67 to 0.71 (for ALL see Table A-8), which were lower than the ones

obtained with 8 replicates (see Figure 4-2), but both numbers of replicates results were

similar in range and behavior. Some relevant differences were a greater difficulty in

achieving convergence and an increased variability in the estimates.

Modeling Global and Local Trends

R-C was the best of all experimental design models fitted assuming independent

errors (ID). This design which incorporates short rows and columns within replicates was

particularly good for patches and in a less extent for gradients. These findings agree with

other studies where R-C designs showed greater efficiency than RCB (Williams et al.

1999) and IB designs (Qiao et al. 2000). Hence, for all surface patterns tested in this

study designing an experiment as R-C is recommended.

Incorporation of the autoregressive error structure yielded improved precision of

treatment estimates as reflected in larger CORR values. The best model, as expected,









corresponded to Full-Poly with AR1AR1+r which was the reference model used to

simulate the surfaces. In contrast, when the autoregressive error structure was fitted

without the nugget effect only small decreases in average CORR values was noted, but

considerable bias in some of the variance components, particularly an underestimation of

the spatial correlation parameters was found. For forestry breeding trials Dutkowski et al.

(2002) also reported an overestimation when the nugget effect was not included in the

model but in their case for the additive genetic variance component. Zimmerman and

Harville (1991) indicated that in many cases this extra parameter might be unnecessary

for most field experiments and could produce problems because of increased

computation. But in this and several other studies, microsite error has been found to be

necessary and often large relative to the total surface error (Ball et al. 1993; Gilmour et

al. 1997; Cullis et al. 1998; Costa e Silva et al. 2001; Dutkowski et al. 2002). All these

results suggest that the inclusion of the nugget or microsite error might be important.

Usually, portions of the gradient variability can be explained by the error structure

(Zimmerman and Harville 1991). In this study, this situation was observed when the

gradient was modeled incorrectly (e.g., using Red-Poly instead of Full-Poly) under an

autoregressive error structure fro GRAD surfaces some improvements on average CORR

values over CR design were found.

There is debate about the use of the design effects together with an error structure

in the same model. These effects induce artificial discontinuities for adjacent

experimental units on the boundaries (replicates or incomplete blocks). In the present

study, once the error structure was modeled through autoregressive with or without

nugget, differences between experimental designs were considerably reduced,









particularly for PATCH surfaces (see Figure 4-2). Nevertheless, in analyses of real

datasets, some authors using autoregressive errors together with design effects have been

successful in increasing precision of treatment effect predictions (Sarker et al. 2001;

Dutkowski et al. 2002).

The concept behind modeling global trends is to represent the field gradient that is

generated by gradual changes in topography and soil characteristics. Polynomial

functions were used in this study, which under independent errors proved to be suitable

only on those simulated surfaces with dominant gradients (ALL and GRAD).

Importantly, once the correct autoregressive error structure was incorporated

(AR1AR1+r) no differences were found between models with different polynomial

functions or design effects. Differences were only noted in the GRAD surface were Full-

Poly was clearly better than any other alternative. In several studies, the incorporation of

fixed or random effects to model global trends together with the autoregressive error

structure has resulted in improved precision of treatment estimates or predictions

(Brownie et al. 1993; Brownie and Gumpertz 1997; Gilmour et al. 1997). Here, trend or

gradient was modeled using polynomial functions, but other options such as smoothing

splines (Verbyla et al. 1999) or differencing the data (Cullis and Gleeson 1991) have

proved useful.

In this study, each simulated dataset was fitted in an automated fashion. Due to the

variability between trials, a general model cannot adequately fit all experiments. For

routine analyses it is recommended, as suggested by Gilmour et al. (1997), that for each

individual datasets several variants of the autoregressive spatial model must be tried, and

the linear model extended by the inclusion of other design factors, polynomial functions,









or other fixed or random effects to model the gradient. Also, model selection should be

always accompanied by the use of diagnostic tools, such as variograms (Cressie 1993) or

the use some of the available mixed model information criteria.

Nearest Neighbor Models

The findings in this study indicated that considerable improvements can be

achieved with PAP or MA methods in controlling for spatial variation. As expected, the

best results were obtained for patchy surfaces. For the same covariate definition, PAP

methods were always better than MA. This situation demonstrates that, as done with

PAP, it is necessary to eliminate the treatment effect from the response variables. The

performance of PAP methods agrees with other studies in which they were always

superior to RCB designs, generally better than IB designs (Kempton and Howes 1981;

Brownie et al. 1993; Vollmann et al. 1996), and only slightly inferior to models that

fitted the autoregressive error structure (Zimmerman and Harville 1991; Brownie et al.

1993; Brownie and Gumpertz 1997).

The best nearest neighbor methods corresponded to PAP-6 and PAP-11, which are

based on 4 and 6 covariates respectively, and yielded better results than the original PAP

method based on a single covariate (PAP-1 and PAP-3). This agrees with other studies

where the inclusion of more than one covariate produced better results (Kempton and

Howes 1981; Vollmann et al. 1996). With extra covariates, it is possible to use more

information about the neighbor plots; therefore, improving efficiency. Nevertheless, for

different situations, the ideal set of covariates needs to be investigated.

PAP methods and other nearest neighbor methods not studied here have the

advantages of simplicity and flexibility. They are simple because only the assumption of

independent errors is required so computational requirements are minimal, and they are









flexible because several covariates with varying number of plots can be defined and

studied. Nevertheless, these methods are not free of difficulties. Testing the inclusion of

covariates in a model is difficult because it is not clear how many degrees of freedom

should be considered for each covariate due to the fact that a previous estimate of the

treatment mean was used. As point of reference, Pearce (1998) in a simulation study

recommended considering 2 degrees of freedom for each covariate. Also, it is not clear

how to treat covariates obtained from border plots, or what to do if no border plots exist.

Ideally, measurements outside the test area should be available as plots that received a

treatment with an average response (Scharf and Alley 1993). Nevertheless, it is common

to use only existing plots to calculate covariates (Wilkinson et al. 1983), or replace non-

existent plots by their interior complement (Magnussen 1993). A similar situation occurs

with missing information that generates an internal border. Another reported difficulty is

that these methods appear to be more sensitive to outliers (Kempton and Howes 1981)

and to the presence of abrupt changes in the error surface (Binns 1987; Pearce 1998); and

in many cases it is not clear what are the ideal solutions for these problems.

The iterative PAP was first suggested by Bartlett (1978) to improve efficiency.

Wilkinson et al. (1983) claimed that there is a considerable positive bias for testing

significance of treatment effects for a single covariate iterative PAP method;

nevertheless, Pearce (1998) in another study found no bias, and showed that some PAP

methods were more effective after iteration. The results in this study for iterated PAP

methods were not favorable producing a reduction in the efficiency as more iterations

were executed. Given the results, non-iterated PAP methods appear to be the best option.

Nevertheless, is important to consider that for the present simulations a relatively large









number of replicates per treatment were available (i.e., 8) that allowed reasonable initial

estimates of treatment means to calculate covariates.

Apparently, the PAP methods can be improved by including model effects to

incorporate gradients or trends (which is where these methods failed to outperform other

options, see Figure 4-4). Design effects, as replicates or blocks have been used in some

studies together with some nearest neighbor covariates (Kempton and Howes 1981;

Scharf and Alley 1993) with varying outcomes. But, as indicated before, the use of

design effects is not recommended because they create artificial discontinuities for

neighboring plots that belong to different blocks. Polynomials functions or smoothing

splines do not have this problem because they produce a continuous gradient; hence, they

are recommended for modeling gradients when employing nearest neighbor methods to

account for patchiness.

Usually, in spatial analysis only positive correlations due to physical proximity are

considered. But, in many real situations, a negative correlation can be generated by

competition between plots belonging to different treatments. This competition can disrupt

the spatial correlation (Kempton and Howes 1981) and might require explicit modeling.

However, this negative correlation is only present once competition for resources (light,

water and nutrients) is important, which might take several years to occur. Because the

simulations in this study did not consider these effects some caution must be taken in

situations with high levels of them.

The use of spatial techniques discussed here does not invalidate classical design of

experiments. As Sarker et al. (2001) suggest, it is important to design the layout of an

experiment according to a preferred design (e.g., IB), which should be first analyzed as









designed, but further spatial models should be fitted to capture other relevant information

(or noise) present in the field. For any design or analysis, treatment randomization is still

critical to neutralize the effect of spatial correlations and provide unbiased estimates and

to validate hypothesis testing (Grondona and Cressie 1991; Brownie et al. 1993). Also,

specific design of experiments in the presence of spatial correlation might yield

improvements, particularly in balancing treatment pairwise comparisons. Little research

in this topic has been reported, and for example, Williams (1985) described an algorithm

to improve IB designs by considering their spatial arrangement.

The results from this study concentrated in single-site analyses. Combining several

sites and other sources of information together can yield to improved estimation of the

genetic parameters. Usually, for multiple-site analyses each test is studied individually to

determine its significant model effects and error structure, and later, all sites are

combined in a unique model using their single-site specifications obtained previously.

Specialized software is required to fit this complex mixed linear model, which commonly

presents difficulties in getting variance components that converge. In contrast, the use of

nearest neighbor models on multiple-site analyses is considerable easier, particularly

because the error structure is simple, and it only requires the definition of site specific

covariate parameters. Nevertheless, further studies in these topics are required.

Conclusions

The main findings of this study indicated that an adequate selection of the

experimental design can produce considerable improvements in the prediction of

treatment effects and better design options than the widely used RCB are available for

field trials in agronomical and forestry studies. Particularly among models assuming

independent errors, R-C designs gave the best results and IB designs of 32 incomplete









blocks were almost as good. Nevertheless, in PATCH surfaces these designs did not

perform as well as autoregressive error structures or nearest neighbor methods. Also,

those designs that modeled global trends (i.e., polynomial functions) performed well in

surfaces with a dominant gradient (GRAD and ALL) but poorly in PATCH.

The incorporation of a separable autoregressive error structure with or without

nugget yielded the highest CORR values among all models and methods tried for all

surface patterns. Here, the inclusion of design effects (replicates and blocks) or

polynomial covariates were only useful when the nugget effect was not included.

Nevertheless, in this last situation biased spatial correlations parameters were found.

Promising results were obtained with the use of nearest neighbor techniques,

particularly for surfaces with patches, but in surfaces with gradients some minor

improvements were also found. PAP methods were always better than MA, and they

were almost as good as methods that modeled the error structure, particularly those

variants that considered more plots and/or covariates. PAP-11 was the best alternative

followed by PAP-6, but the use of the latter is recommended because of its simplicity.

The use of an iterated PAP method did not produce improvements in this study,

generating deteriorated estimates.

In summary, if simple analyses are preferred, R-C and IB designs with

independent errors should be preferred. For further improvements, it is recommended that

some form of PAP methods be used, but several covariates must be tried and tested.

Finally, spatial analysis incorporating error structures are promising, and they should be

used whenever possible, but the computational requirements and some uncertainty about

the correct procedures and testing may limit its practical use.














CHAPTER 5
POST-HOC BLOCKING TO IMPROVE HERITABILITY AND
BREEDING VALUE PREDICTION

Introduction

Genetic testing of provenances, varieties and families in forestry and agricultural

crops is usually a long process and frequently one of the most expensive activities of any

genetic improvement program (Zobel and Talbert 1984, p. 232). Appropriate selection of

an experimental design and statistical analysis can produce considerable improvement in

the prediction of breeding values and efficient use of available resources.

Traditionally, in forest trials the randomized complete block (RCB) design has

been favored. RCB design is effective when within replicate (or block) variability is

relatively small, a situation that is rare in forest sites, or occurs only with relatively small

replicates (Costa e Silva et al. 2001). Since a large number of genetic entities are usually

tested together, a good alternative is to implement incomplete block designs (IB) which

subdivide the full replicate into smaller compartments (i.e., incomplete blocks) that tend

to be more homogeneous. For these designs, several authors have reported greater

efficiencies in comparison to RCB (Fu et al. 1998; Fu et al. 1999; see Chapter 2). It is

also possible to simultaneously implement two-way blocking using row-column designs

(R-C). These R-C designs are described in detail by John and Williams (1995, p. 87) and

often yield even greater efficiencies than IB designs (Lin et al. 1993; Qiao et al. 2000;

see also Chapter 2).









Fitting the best linear model that adequately describes the trial data together with

the appropriate specification of random and fixed effects is critical. In the literature

several techniques are available for improving statistical analyses, such as 1) spatial

models (Gilmour et al. 1997; Chapter 4); 2) nearest neighbor methods (Vollmann et al.

1996; Chapter 4); and 3) inclusion of covariates and other factors to deal with missing

values (Cochran 1957) or nuisance effects (Gilmour et al. 1997). These are examples of

"a posteriori" analyses. In these options, an extended linear model different than the

original experimental design is fitted, which aims to model elements that were unnoticed

or not controlled at test establishment.

Another "a posteriori" technique, called post-hoc blocking, can also be easily

implemented. This method consists of superimposing a blocking structure on top of the

original field design and a linear model is fitted as if the blocking effects were present in

the original design. Usually, an IB design is fitted over a simpler design (e.g., RCB), but

an R-C can also be superimposed. This technique was originally proposed by Patterson

and Hunter (1983) as a tool to inexpensively evaluate the efficiency of potential

experimental designs. Nevertheless, several authors have used post-hoc blocking

successfully to increase heritability and the precision of prediction of genetic effects for

final analyses (Ericsson 1997; Dutkowski et al. 2002; Lopez et al. 2002).

The primary problem in implementing post-hoc blocking is determining the

characteristics of the incomplete blocks to be superimposed. If large incomplete blocks

are used, the blocks will continue to be highly heterogeneous. On the other hand, small

incomplete blocks should considerably reduce the residual variability by capturing

smaller areas (or patches) of the environmental surface. But if block size becomes too









small, fewer treatments will occur together in the same block; therefore, the standard

error of their difference can be considerably larger. Also, according to Ericsson (1997),

with small blocks part of the treatment (or genetic) variability can be absorbed by the

block variance due to an increased chance that two good (or bad) genotypes occur

together in the same block. Another problem is the shape of the incomplete block;

nevertheless, general guidelines used for design of experiments can be followed. For

example, a shape as close to a square as possible, and orienting the blocks main axis

perpendicularly to the principal field gradient should be preferred (Zobel and Talbert

1984, p. 248).

The present study used simulated single-site clonal trials and 3 different

environmental patterns to understand consequences of and define future strategies for

using post-hoc blocking in estimating heritabilities and predicting breeding values.

Specifically, the objectives were to 1) study the use of post-hoc blocking for several

experimental designs, surface patterns and numbers of ramets; 2) verify a potential

reduction in the genetic variance as the incomplete blocks become smaller; 3) compare

the performance of different strategies or criteria for selecting an appropriate blocking

structure; and 4) discuss statistical and practical issues related to implementing post-hoc

blocking.

Materials and Methods

Several single site clonal trials were simulated based on a rectangular grid of 64

rows and 32 columns with square spacing and no missing observations (see Chapter 2 for

details). Briefly, 256 clones each with 8 ramets were "planted" in single-tree plots

according to several experimental plans. The simulated environmental or surface patterns

were based on two main elements: a gradient or trend generated with a third degree









continuous polynomial function, and patches that were incorporated through a covariance

structure based in a first-order separable autoregressive process or AR1AR1 (Cressie

1993; Gilmour et al. 1997). The 3 surface patterns considered were 1) only patches

(PATCH); 2) only gradients (GRAD); and 3) both components (ALL). The variance

structure was set for a completely randomized design to have a single-site individual

broad-sense heritability (Hi2) of 0.25 originating from a total variance fixed at 1.0 and a

clonal variance of 2clione = 0.25. The simulated designs were a randomized complete

design with 8 resolvable replicates (RCB), incomplete block designs with 4, 8, 16 and 32

incomplete blocks within each of the 8 full replicates (IB 4, IB 8, IB 16 and IB 32), and a

row-column design with 16 rows and 16 columns (R-C). For each design, several

implementations were generated using the software CycDesigN (Whitaker et al. 2002)

that produces a-designs. For each combination of surface pattern and design type 1,000

different datasets were generated.

The software ASREML (Gilmour et al. 2002) was used to fit all mixed linear

models and to obtain restricted maximum likelihood (REML) variance component

estimates and best linear unbiased predictions (BLUP) of clonal values (Patterson and

Thompson 1971). To compare alternatives, several statistics were calculated: summary

statistics of estimated variance components, average empirical correlation between true

and predicted clonal BLUP values (CORR) and individual single-site heritabilities

(H2 ). Finally, for each linear model and dataset the log-likelihood value (logL) was

recorded together with its average standard error of the difference (SED) between any

two clones.







81


Table 5-1. Linear models fitted for simulated datasets using classical experimental
designs to model global trends over all surface patterns: randomized complete
block (RCB), incomplete block design with x blocks (IB), and a row-column
(R-C) design. All model effects other than the mean were considered random.

Model a
RCB y, = + rep, + clone, + r,
IB x Yk = + rep1 + block(rep), + clonek + Ek
R-C Yljk = + rep, + col(rep), + row(rep)1k + clone1 + skl
a clone, clone effect; rep, resolvable replication; block(rep), incomplete block nested within replication;
col(rep), column nested within replication (short column); row(rep), row nested within replication (short
row); E, residual.


IB4-8x8


B 8-8x4


TB 16-4x4

1 5 9 13

2 6 10 14

3 7 11 15

4 8 12 16


IB 32-4x2


IB64-2x2
I1 I1 1 1 I


IB 128 2 x 1


Figure 5-1. Experimental layout for randomized complete block design simulations with
8 resolvable replicates (left) and partitioning of one replicate for incomplete
block design layouts (right). All simulations had 256 unrelated clones with 8
ramets per clone per site for a total of 2,048 trees.


32 trees
_^___________>


161


16 trees









The first stage in this study consisted of fitting the original experimental designs

(without post-hoc blocking) using the corresponding linear models (see Table 5-1). Later,

by employing only the RCB designs for all surface patterns, post-hoc blocking was

implemented by superimposing 4, 8, 16 and 32 incomplete blocks of sizes 8 x 8, 8 x 4, 4

x 4, and 4 trees x 2 trees, respectively, and a R-C design with short rows and columns of

length 16 (see Figure 5-1). These "new" post-hoc designs were fitted using the

corresponding linear model from Table 5-1. Statistical comparisons of individual

heritabilities between original designs and post-hoc blocking were made using a z-test.

In a second stage, the first 500 datasets generated for RCB designs were used to

perform additional post-hoc blocking studies. The same post-hoc blocking designs

previously used were fitted and two extra incomplete block designs were incorporated to

examine the effects of very small incomplete blocks: a layout with 64 (IB 64) and 128

(IB 128) blocks were superimposed with block sizes of 2 x 2 and 2 x 1 respectively (see

Figure 5-1). The linear models from Table 5-1 were fitted to all datasets, and the

previously described statistics were also calculated. Additionally, for each of the 500

datasets, a subset of 2 contiguous replicates was selected at random constituting a "new"

trial based on only 2 ramets per clone. This reduced trial size was used because of interest

in studying the potential effects of using post-hoc blocking in the estimation of the

variance components and other statistics for trials with fewer replications. For these

datasets, the same incomplete block and row-column designs described previously were

fitted (IB 4, IB 8, IB 16, IB 32, IB 64, IB 128 and R-C), and summary statistics were

calculated.









In practical situations, several blocking alternatives (varying numbers of blocks and

block shapes) are tested during analysis, and a final model must be selected. In this study,

the performance of log-likelihood and SED decision criteria was studied and compared

with individual heritabilities. The first criteria, logL, selects the blocking structure that

maximizes the log-likelihood of the mixed linear model (Lopez et al. 2002). For SED, the

model with the smallest average standard deviation of the difference between treatments

is selected. The procedure consisted of selecting and recording the best model fitted in

the second stage using both of these criteria for each of the 500 datasets and experimental

designs. Here, frequency tables were obtained with all designs together, and only for IB

designs.

Results

When results from fitting the original experimental designs were compared to post-

hoc blocking designs of the same block size and orientation, small differences were found

for all statistics. The largest difference for average HR was 0.003 units (Table A-9).

Further, z-tests comparing individual heritabilities between the original design and post-

hoc blocking for each design type and surface pattern indicated that no significant

differences existed (experiment-wise a = 0.05), with the single exception of R-C designs

on PATCH surfaces. Also, when standard deviations for HE2 were compared between

original and post-hoc blocking analyses, the differences in magnitude were very small

and probably due to sampling variation (Table A-9). The average clonal variance did not

differ between original and post-hoc blocking, and a negligible decrease in the precision

of this component was found for post-hoc blocking. A similar situation of very little










difference between original experimental designs and post-hoc blocking was also

observed for average correlation between true and predicted clonal values (Figure 5-2).

Average variance components obtained in the second stage employing even smaller

post-hoc blocks for all designs and surface patterns for the case of 8 ramets per clone are

shown in Table 5-2. The clonal variance was always close to its parametric value of 0.25,

with no reduction in its average value as the size of the block decreased, as postulated by

Ericsson (1997). Also, the block variance increased and the error variance decreased as

the number of incomplete blocks increased. Considerable differences were noted in block

variance components between IB 4 and IB 128, indicating a large influence of the

experimental design used on partitioning total variance.


0.90



0 .88 -. ... .. .... ........ .





0.86 -
0.86



ALL -Original
**0-- ALL Post-hoc
0 4 -- PATCH Original
.-A-. PATCH- Post-hoc
GRAD Original
*-V-- GRAD- Post-hoc

0.82 1 1
RCB IB4 IB8 IB16 IB32 R-C

Figure 5-2. Average correlation between true and predicted clonal values (CORR) for
original and post-hoc blocking analyses on ALL, PATCH and GRAD surface
patterns for simulated surfaces with 8 ramets per clone.










Table 5-2. Average variance components and individual broad-sense heritability for all
surface patterns obtained from fitting the models in Table 5-1 from simulated
datasets with a randomized complete block designs with 8 ramets per clone
per site.

ALL a
Model b clone rep block(rep) col(rep) row(rep) error Hi
RCB 0.25 0.17 0.60 0.296
IB 4 0.25 0.16 0.05 0.56 0.310
IB 8 0.25 0.17 0.07 0.54 0.319
IB 16 0.25 0.17 0.08 0.53 0.323
IB 32 0.25 0.17 0.09 0.51 0.329
IB 64 0.25 0.17 0.10 0.50 0.337
IB 128 0.25 0.17 0.10 0.50 0.336
R-C 0.25 0.17 0.06 0.05 0.50 0.336

PATCH
Model clone rep block(rep) col(rep) row(rep) error HE2
RCB 0.25 0.02 0.73 0.256
IB 4 0.25 0.01 0.03 0.71 0.263
IB 8 0.25 0.01 0.05 0.68 0.269
IB 16 0.25 0.02 0.07 0.67 0.274
IB 32 0.25 0.02 0.10 0.63 0.284
IB 64 0.25 0.02 0.13 0.60 0.295
IB 128 0.25 0.02 0.15 0.58 0.304
R-C 0.25 0.01 0.06 0.06 0.62 0.288

GRAD
Model clone rep block(rep) col(rep) row(rep) error HE,
RCB 0.25 0.17 0.61 0.292
IB 4 0.25 0.15 0.05 0.57 0.305
IB 8 0.25 0.16 0.05 0.56 0.308
IB 16 0.25 0.16 0.05 0.56 0.308
IB 32 0.25 0.16 0.05 0.56 0.308
IB 64 0.25 0.16 0.06 0.54 0.314
IB 128 0.25 0.17 0.06 0.55 0.313
R-C 0.25 0.16 0.04 0.02 0.55 0.312
a ALL, surfaces with patches and gradients; GRAD, surfaces with only gradients; PATCH, surfaces with
only patches.
b CR, complete randomized; RCB, randomized complete block; IB x, incomplete block with x blocks per
replicate; R-C, row-column.
c rep, resolvable replicate; block(rep), incomplete block nested within replicate; col(rep), column nested
within replicate; row(rep), row nested within replicate; clone, clone; e, residual.







86




0.34 -- -- -- -- -- -- -- 0.90
a- b


0.32 0.89-



5 0.30 / 0.88- V I

I *o O
b 0
b 0.28 0.87-
SO..0
0o'' ..0.0.

.0 I ---- I0 "
0.26 -0- ALL 0.86- O
0 ..O0... PATCH
--- GRAD O

0.24 0.85




Figure 5-3. Average individual heritability (a) and correlation between true and predicted
clonal values (b) calculated over 500 simulations with a randomized complete
block designs with 8 ramets per clone per site for each surface pattern and
post-hoc design.



For average individual heritability (Figure 5-3a) IB 128 was the best design across

all 3 surface patterns, but IB 64 and R-C were relatively close. In fact, considerable


improvements in H,2 can be achieved with post-hoc blocking using better designs;


average H values of up to 0.34, compared with a base value of 0.25 for the completely


randomized design were obtained for smaller block sizes. Nevertheless, a better statistic

to use is the average correlation between the true and predicted values (CORR) because it

evaluates how well a post-hoc blocking design predicts clonal breeding values and,


hence, genetic gain from clonal selection. The trends between H,2 and CORR differ


considerably (see Figure 5-3), and a decrease in CORR was present for those designs